lnet/lnet/lib-move.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2012, 2017, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  *
  31  * lnet/lnet/lib-move.c
  32  *
  33  * Data movement routines
  34  */
  35
  36 #define DEBUG_SUBSYSTEM S_LNET
  37
  38 #include <linux/pagemap.h>
  39
  40 #include <lnet/lib-lnet.h>
  41 #include <linux/nsproxy.h>
  42 #include <lnet/lnet_rdma.h>
  43 #include <net/net_namespace.h>
  44
  45 static int local_nid_dist_zero = 1;
  46 module_param(local_nid_dist_zero, int, 0444);
  47 MODULE_PARM_DESC(local_nid_dist_zero, "Reserved");
  48
  49 struct lnet_send_data {
  50         struct lnet_ni *sd_best_ni;
  51         struct lnet_peer_ni *sd_best_lpni;
  52         struct lnet_peer_ni *sd_final_dst_lpni;
  53         struct lnet_peer *sd_peer;
  54         struct lnet_peer *sd_gw_peer;
  55         struct lnet_peer_ni *sd_gw_lpni;
  56         struct lnet_peer_net *sd_peer_net;
  57         struct lnet_msg *sd_msg;
  58         lnet_nid_t sd_dst_nid;
  59         lnet_nid_t sd_src_nid;
  60         lnet_nid_t sd_rtr_nid;
  61         int sd_cpt;
  62         int sd_md_cpt;
  63         __u32 sd_send_case;
  64 };
  65
  66 static inline bool
  67 lnet_msg_is_response(struct lnet_msg *msg)
  68 {
  69         return msg->msg_type == LNET_MSG_ACK || msg->msg_type == LNET_MSG_REPLY;
  70 }
  71
  72 static inline bool
  73 lnet_response_tracking_enabled(__u32 msg_type, unsigned int md_options)
  74 {
  75         if (md_options & LNET_MD_NO_TRACK_RESPONSE)
  76                 /* Explicitly disabled in MD options */
  77                 return false;
  78
  79         if (md_options & LNET_MD_TRACK_RESPONSE)
  80                 /* Explicity enabled in MD options */
  81                 return true;
  82
  83         if (lnet_response_tracking == 3)
  84                 /* Enabled for all message types */
  85                 return true;
  86
  87         if (msg_type == LNET_MSG_PUT)
  88                 return lnet_response_tracking == 2;
  89
  90         if (msg_type == LNET_MSG_GET)
  91                 return lnet_response_tracking == 1;
  92
  93         return false;
  94 }
  95
  96 static inline struct lnet_comm_count *
  97 get_stats_counts(struct lnet_element_stats *stats,
  98                  enum lnet_stats_type stats_type)
  99 {
 100         switch (stats_type) {
 101         case LNET_STATS_TYPE_SEND:
 102                 return &stats->el_send_stats;
 103         case LNET_STATS_TYPE_RECV:
 104                 return &stats->el_recv_stats;
 105         case LNET_STATS_TYPE_DROP:
 106                 return &stats->el_drop_stats;
 107         default:
 108                 CERROR("Unknown stats type\n");
 109         }
 110
 111         return NULL;
 112 }
 113
 114 void lnet_incr_stats(struct lnet_element_stats *stats,
 115                      enum lnet_msg_type msg_type,
 116                      enum lnet_stats_type stats_type)
 117 {
 118         struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
 119         if (!counts)
 120                 return;
 121
 122         switch (msg_type) {
 123         case LNET_MSG_ACK:
 124                 atomic_inc(&counts->co_ack_count);
 125                 break;
 126         case LNET_MSG_PUT:
 127                 atomic_inc(&counts->co_put_count);
 128                 break;
 129         case LNET_MSG_GET:
 130                 atomic_inc(&counts->co_get_count);
 131                 break;
 132         case LNET_MSG_REPLY:
 133                 atomic_inc(&counts->co_reply_count);
 134                 break;
 135         case LNET_MSG_HELLO:
 136                 atomic_inc(&counts->co_hello_count);
 137                 break;
 138         default:
 139                 CERROR("There is a BUG in the code. Unknown message type\n");
 140                 break;
 141         }
 142 }
 143
 144 __u32 lnet_sum_stats(struct lnet_element_stats *stats,
 145                      enum lnet_stats_type stats_type)
 146 {
 147         struct lnet_comm_count *counts = get_stats_counts(stats, stats_type);
 148         if (!counts)
 149                 return 0;
 150
 151         return (atomic_read(&counts->co_ack_count) +
 152                 atomic_read(&counts->co_put_count) +
 153                 atomic_read(&counts->co_get_count) +
 154                 atomic_read(&counts->co_reply_count) +
 155                 atomic_read(&counts->co_hello_count));
 156 }
 157
 158 static inline void assign_stats(struct lnet_ioctl_comm_count *msg_stats,
 159                                 struct lnet_comm_count *counts)
 160 {
 161         msg_stats->ico_get_count = atomic_read(&counts->co_get_count);
 162         msg_stats->ico_put_count = atomic_read(&counts->co_put_count);
 163         msg_stats->ico_reply_count = atomic_read(&counts->co_reply_count);
 164         msg_stats->ico_ack_count = atomic_read(&counts->co_ack_count);
 165         msg_stats->ico_hello_count = atomic_read(&counts->co_hello_count);
 166 }
 167
 168 void lnet_usr_translate_stats(struct lnet_ioctl_element_msg_stats *msg_stats,
 169                               struct lnet_element_stats *stats)
 170 {
 171         struct lnet_comm_count *counts;
 172
 173         LASSERT(msg_stats);
 174         LASSERT(stats);
 175
 176         counts = get_stats_counts(stats, LNET_STATS_TYPE_SEND);
 177         if (!counts)
 178                 return;
 179         assign_stats(&msg_stats->im_send_stats, counts);
 180
 181         counts = get_stats_counts(stats, LNET_STATS_TYPE_RECV);
 182         if (!counts)
 183                 return;
 184         assign_stats(&msg_stats->im_recv_stats, counts);
 185
 186         counts = get_stats_counts(stats, LNET_STATS_TYPE_DROP);
 187         if (!counts)
 188                 return;
 189         assign_stats(&msg_stats->im_drop_stats, counts);
 190 }
 191
 192 int
 193 lnet_fail_nid(lnet_nid_t nid4, unsigned int threshold)
 194 {
 195         struct lnet_test_peer *tp;
 196         struct list_head *el;
 197         struct list_head *next;
 198         struct lnet_nid nid;
 199         LIST_HEAD(cull);
 200
 201         lnet_nid4_to_nid(nid4, &nid);
 202         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
 203         if (threshold != 0) {
 204                 /* Adding a new entry */
 205                 LIBCFS_ALLOC(tp, sizeof(*tp));
 206                 if (tp == NULL)
 207                         return -ENOMEM;
 208
 209                 tp->tp_nid = nid;
 210                 tp->tp_threshold = threshold;
 211
 212                 lnet_net_lock(0);
 213                 list_add_tail(&tp->tp_list, &the_lnet.ln_test_peers);
 214                 lnet_net_unlock(0);
 215                 return 0;
 216         }
 217
 218         lnet_net_lock(0);
 219
 220         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
 221                 tp = list_entry(el, struct lnet_test_peer, tp_list);
 222
 223                 if (tp->tp_threshold == 0 ||    /* needs culling anyway */
 224                     LNET_NID_IS_ANY(&nid) ||    /* removing all entries */
 225                     nid_same(&tp->tp_nid, &nid)) {      /* matched this one */
 226                         list_move(&tp->tp_list, &cull);
 227                 }
 228         }
 229
 230         lnet_net_unlock(0);
 231
 232         while (!list_empty(&cull)) {
 233                 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
 234
 235                 list_del(&tp->tp_list);
 236                 LIBCFS_FREE(tp, sizeof(*tp));
 237         }
 238         return 0;
 239 }
 240
 241 static int
 242 fail_peer(lnet_nid_t nid4, int outgoing)
 243 {
 244         struct lnet_test_peer *tp;
 245         struct list_head *el;
 246         struct list_head *next;
 247         struct lnet_nid nid;
 248         LIST_HEAD(cull);
 249         int fail = 0;
 250
 251         lnet_nid4_to_nid(nid4, &nid);
 252         /* NB: use lnet_net_lock(0) to serialize operations on test peers */
 253         lnet_net_lock(0);
 254
 255         list_for_each_safe(el, next, &the_lnet.ln_test_peers) {
 256                 tp = list_entry(el, struct lnet_test_peer, tp_list);
 257
 258                 if (tp->tp_threshold == 0) {
 259                         /* zombie entry */
 260                         if (outgoing) {
 261                                 /* only cull zombies on outgoing tests,
 262                                  * since we may be at interrupt priority on
 263                                  * incoming messages. */
 264                                 list_move(&tp->tp_list, &cull);
 265                         }
 266                         continue;
 267                 }
 268
 269                 if (LNET_NID_IS_ANY(&tp->tp_nid) ||     /* fail every peer */
 270                     nid_same(&nid, &tp->tp_nid)) {      /* fail this peer */
 271                         fail = 1;
 272
 273                         if (tp->tp_threshold != LNET_MD_THRESH_INF) {
 274                                 tp->tp_threshold--;
 275                                 if (outgoing &&
 276                                     tp->tp_threshold == 0) {
 277                                         /* see above */
 278                                         list_move(&tp->tp_list, &cull);
 279                                 }
 280                         }
 281                         break;
 282                 }
 283         }
 284
 285         lnet_net_unlock(0);
 286
 287         while (!list_empty(&cull)) {
 288                 tp = list_entry(cull.next, struct lnet_test_peer, tp_list);
 289                 list_del(&tp->tp_list);
 290
 291                 LIBCFS_FREE(tp, sizeof(*tp));
 292         }
 293
 294         return fail;
 295 }
 296
 297 unsigned int
 298 lnet_iov_nob(unsigned int niov, struct kvec *iov)
 299 {
 300         unsigned int nob = 0;
 301
 302         LASSERT(niov == 0 || iov != NULL);
 303         while (niov-- > 0)
 304                 nob += (iov++)->iov_len;
 305
 306         return (nob);
 307 }
 308 EXPORT_SYMBOL(lnet_iov_nob);
 309
 310 void
 311 lnet_copy_iov2iov(unsigned int ndiov, struct kvec *diov, unsigned int doffset,
 312                   unsigned int nsiov, struct kvec *siov, unsigned int soffset,
 313                   unsigned int nob)
 314 {
 315         /* NB diov, siov are READ-ONLY */
 316         unsigned int this_nob;
 317
 318         if (nob == 0)
 319                 return;
 320
 321         /* skip complete frags before 'doffset' */
 322         LASSERT(ndiov > 0);
 323         while (doffset >= diov->iov_len) {
 324                 doffset -= diov->iov_len;
 325                 diov++;
 326                 ndiov--;
 327                 LASSERT(ndiov > 0);
 328         }
 329
 330         /* skip complete frags before 'soffset' */
 331         LASSERT(nsiov > 0);
 332         while (soffset >= siov->iov_len) {
 333                 soffset -= siov->iov_len;
 334                 siov++;
 335                 nsiov--;
 336                 LASSERT(nsiov > 0);
 337         }
 338
 339         do {
 340                 LASSERT(ndiov > 0);
 341                 LASSERT(nsiov > 0);
 342                 this_nob = min3((unsigned int)diov->iov_len - doffset,
 343                                 (unsigned int)siov->iov_len - soffset,
 344                                 nob);
 345
 346                 memcpy((char *)diov->iov_base + doffset,
 347                        (char *)siov->iov_base + soffset, this_nob);
 348                 nob -= this_nob;
 349
 350                 if (diov->iov_len > doffset + this_nob) {
 351                         doffset += this_nob;
 352                 } else {
 353                         diov++;
 354                         ndiov--;
 355                         doffset = 0;
 356                 }
 357
 358                 if (siov->iov_len > soffset + this_nob) {
 359                         soffset += this_nob;
 360                 } else {
 361                         siov++;
 362                         nsiov--;
 363                         soffset = 0;
 364                 }
 365         } while (nob > 0);
 366 }
 367 EXPORT_SYMBOL(lnet_copy_iov2iov);
 368
 369 unsigned int
 370 lnet_kiov_nob(unsigned int niov, struct bio_vec *kiov)
 371 {
 372         unsigned int  nob = 0;
 373
 374         LASSERT(niov == 0 || kiov != NULL);
 375         while (niov-- > 0)
 376                 nob += (kiov++)->bv_len;
 377
 378         return (nob);
 379 }
 380 EXPORT_SYMBOL(lnet_kiov_nob);
 381
 382 void
 383 lnet_copy_kiov2kiov(unsigned int ndiov, struct bio_vec *diov,
 384                     unsigned int doffset,
 385                     unsigned int nsiov, struct bio_vec *siov,
 386                     unsigned int soffset,
 387                     unsigned int nob)
 388 {
 389         /* NB diov, siov are READ-ONLY */
 390         unsigned int    this_nob;
 391         char           *daddr = NULL;
 392         char           *saddr = NULL;
 393
 394         if (nob == 0)
 395                 return;
 396
 397         LASSERT (!in_interrupt ());
 398
 399         LASSERT (ndiov > 0);
 400         while (doffset >= diov->bv_len) {
 401                 doffset -= diov->bv_len;
 402                 diov++;
 403                 ndiov--;
 404                 LASSERT(ndiov > 0);
 405         }
 406
 407         LASSERT(nsiov > 0);
 408         while (soffset >= siov->bv_len) {
 409                 soffset -= siov->bv_len;
 410                 siov++;
 411                 nsiov--;
 412                 LASSERT(nsiov > 0);
 413         }
 414
 415         do {
 416                 LASSERT(ndiov > 0);
 417                 LASSERT(nsiov > 0);
 418                 this_nob = min3(diov->bv_len - doffset,
 419                                 siov->bv_len - soffset,
 420                                 nob);
 421
 422                 if (daddr == NULL)
 423                         daddr = ((char *)kmap(diov->bv_page)) +
 424                                 diov->bv_offset + doffset;
 425                 if (saddr == NULL)
 426                         saddr = ((char *)kmap(siov->bv_page)) +
 427                                 siov->bv_offset + soffset;
 428
 429                 /* Vanishing risk of kmap deadlock when mapping 2 pages.
 430                  * However in practice at least one of the kiovs will be mapped
 431                  * kernel pages and the map/unmap will be NOOPs */
 432
 433                 memcpy (daddr, saddr, this_nob);
 434                 nob -= this_nob;
 435
 436                 if (diov->bv_len > doffset + this_nob) {
 437                         daddr += this_nob;
 438                         doffset += this_nob;
 439                 } else {
 440                         kunmap(diov->bv_page);
 441                         daddr = NULL;
 442                         diov++;
 443                         ndiov--;
 444                         doffset = 0;
 445                 }
 446
 447                 if (siov->bv_len > soffset + this_nob) {
 448                         saddr += this_nob;
 449                         soffset += this_nob;
 450                 } else {
 451                         kunmap(siov->bv_page);
 452                         saddr = NULL;
 453                         siov++;
 454                         nsiov--;
 455                         soffset = 0;
 456                 }
 457         } while (nob > 0);
 458
 459         if (daddr != NULL)
 460                 kunmap(diov->bv_page);
 461         if (saddr != NULL)
 462                 kunmap(siov->bv_page);
 463 }
 464 EXPORT_SYMBOL(lnet_copy_kiov2kiov);
 465
 466 void
 467 lnet_copy_kiov2iov (unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 468                     unsigned int nkiov, struct bio_vec *kiov,
 469                     unsigned int kiovoffset,
 470                     unsigned int nob)
 471 {
 472         /* NB iov, kiov are READ-ONLY */
 473         unsigned int    this_nob;
 474         char           *addr = NULL;
 475
 476         if (nob == 0)
 477                 return;
 478
 479         LASSERT (!in_interrupt ());
 480
 481         LASSERT (niov > 0);
 482         while (iovoffset >= iov->iov_len) {
 483                 iovoffset -= iov->iov_len;
 484                 iov++;
 485                 niov--;
 486                 LASSERT(niov > 0);
 487         }
 488
 489         LASSERT(nkiov > 0);
 490         while (kiovoffset >= kiov->bv_len) {
 491                 kiovoffset -= kiov->bv_len;
 492                 kiov++;
 493                 nkiov--;
 494                 LASSERT(nkiov > 0);
 495         }
 496
 497         do {
 498                 LASSERT(niov > 0);
 499                 LASSERT(nkiov > 0);
 500                 this_nob = min3((unsigned int)iov->iov_len - iovoffset,
 501                                 (unsigned int)kiov->bv_len - kiovoffset,
 502                                 nob);
 503
 504                 if (addr == NULL)
 505                         addr = ((char *)kmap(kiov->bv_page)) +
 506                                 kiov->bv_offset + kiovoffset;
 507
 508                 memcpy((char *)iov->iov_base + iovoffset, addr, this_nob);
 509                 nob -= this_nob;
 510
 511                 if (iov->iov_len > iovoffset + this_nob) {
 512                         iovoffset += this_nob;
 513                 } else {
 514                         iov++;
 515                         niov--;
 516                         iovoffset = 0;
 517                 }
 518
 519                 if (kiov->bv_len > kiovoffset + this_nob) {
 520                         addr += this_nob;
 521                         kiovoffset += this_nob;
 522                 } else {
 523                         kunmap(kiov->bv_page);
 524                         addr = NULL;
 525                         kiov++;
 526                         nkiov--;
 527                         kiovoffset = 0;
 528                 }
 529
 530         } while (nob > 0);
 531
 532         if (addr != NULL)
 533                 kunmap(kiov->bv_page);
 534 }
 535 EXPORT_SYMBOL(lnet_copy_kiov2iov);
 536
 537 void
 538 lnet_copy_iov2kiov(unsigned int nkiov, struct bio_vec *kiov,
 539                    unsigned int kiovoffset,
 540                    unsigned int niov, struct kvec *iov, unsigned int iovoffset,
 541                    unsigned int nob)
 542 {
 543         /* NB kiov, iov are READ-ONLY */
 544         unsigned int    this_nob;
 545         char           *addr = NULL;
 546
 547         if (nob == 0)
 548                 return;
 549
 550         LASSERT (!in_interrupt ());
 551
 552         LASSERT (nkiov > 0);
 553         while (kiovoffset >= kiov->bv_len) {
 554                 kiovoffset -= kiov->bv_len;
 555                 kiov++;
 556                 nkiov--;
 557                 LASSERT(nkiov > 0);
 558         }
 559
 560         LASSERT(niov > 0);
 561         while (iovoffset >= iov->iov_len) {
 562                 iovoffset -= iov->iov_len;
 563                 iov++;
 564                 niov--;
 565                 LASSERT(niov > 0);
 566         }
 567
 568         do {
 569                 LASSERT(nkiov > 0);
 570                 LASSERT(niov > 0);
 571                 this_nob = min3((unsigned int)kiov->bv_len - kiovoffset,
 572                                 (unsigned int)iov->iov_len - iovoffset,
 573                                 nob);
 574
 575                 if (addr == NULL)
 576                         addr = ((char *)kmap(kiov->bv_page)) +
 577                                 kiov->bv_offset + kiovoffset;
 578
 579                 memcpy (addr, (char *)iov->iov_base + iovoffset, this_nob);
 580                 nob -= this_nob;
 581
 582                 if (kiov->bv_len > kiovoffset + this_nob) {
 583                         addr += this_nob;
 584                         kiovoffset += this_nob;
 585                 } else {
 586                         kunmap(kiov->bv_page);
 587                         addr = NULL;
 588                         kiov++;
 589                         nkiov--;
 590                         kiovoffset = 0;
 591                 }
 592
 593                 if (iov->iov_len > iovoffset + this_nob) {
 594                         iovoffset += this_nob;
 595                 } else {
 596                         iov++;
 597                         niov--;
 598                         iovoffset = 0;
 599                 }
 600         } while (nob > 0);
 601
 602         if (addr != NULL)
 603                 kunmap(kiov->bv_page);
 604 }
 605 EXPORT_SYMBOL(lnet_copy_iov2kiov);
 606
 607 int
 608 lnet_extract_kiov(int dst_niov, struct bio_vec *dst,
 609                   int src_niov, struct bio_vec *src,
 610                   unsigned int offset, unsigned int len)
 611 {
 612         /* Initialise 'dst' to the subset of 'src' starting at 'offset',
 613          * for exactly 'len' bytes, and return the number of entries.
 614          * NB not destructive to 'src' */
 615         unsigned int    frag_len;
 616         unsigned int    niov;
 617
 618         if (len == 0)                           /* no data => */
 619                 return (0);                     /* no frags */
 620
 621         LASSERT(src_niov > 0);
 622         while (offset >= src->bv_len) {      /* skip initial frags */
 623                 offset -= src->bv_len;
 624                 src_niov--;
 625                 src++;
 626                 LASSERT(src_niov > 0);
 627         }
 628
 629         niov = 1;
 630         for (;;) {
 631                 LASSERT(src_niov > 0);
 632                 LASSERT((int)niov <= dst_niov);
 633
 634                 frag_len = src->bv_len - offset;
 635                 dst->bv_page = src->bv_page;
 636                 dst->bv_offset = src->bv_offset + offset;
 637
 638                 if (len <= frag_len) {
 639                         dst->bv_len = len;
 640                         LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
 641                         return niov;
 642                 }
 643
 644                 dst->bv_len = frag_len;
 645                 LASSERT(dst->bv_offset + dst->bv_len <= PAGE_SIZE);
 646
 647                 len -= frag_len;
 648                 dst++;
 649                 src++;
 650                 niov++;
 651                 src_niov--;
 652                 offset = 0;
 653         }
 654 }
 655 EXPORT_SYMBOL(lnet_extract_kiov);
 656
 657 void
 658 lnet_ni_recv(struct lnet_ni *ni, void *private, struct lnet_msg *msg,
 659              int delayed, unsigned int offset, unsigned int mlen,
 660              unsigned int rlen)
 661 {
 662         unsigned int niov = 0;
 663         struct kvec *iov = NULL;
 664         struct bio_vec  *kiov = NULL;
 665         int rc;
 666
 667         LASSERT (!in_interrupt ());
 668         LASSERT (mlen == 0 || msg != NULL);
 669
 670         if (msg != NULL) {
 671                 LASSERT(msg->msg_receiving);
 672                 LASSERT(!msg->msg_sending);
 673                 LASSERT(rlen == msg->msg_len);
 674                 LASSERT(mlen <= msg->msg_len);
 675                 LASSERT(msg->msg_offset == offset);
 676                 LASSERT(msg->msg_wanted == mlen);
 677
 678                 msg->msg_receiving = 0;
 679
 680                 if (mlen != 0) {
 681                         niov = msg->msg_niov;
 682                         kiov = msg->msg_kiov;
 683
 684                         LASSERT (niov > 0);
 685                         LASSERT ((iov == NULL) != (kiov == NULL));
 686                 }
 687         }
 688
 689         rc = (ni->ni_net->net_lnd->lnd_recv)(ni, private, msg, delayed,
 690                                              niov, kiov, offset, mlen,
 691                                              rlen);
 692         if (rc < 0)
 693                 lnet_finalize(msg, rc);
 694 }
 695
 696 static void
 697 lnet_setpayloadbuffer(struct lnet_msg *msg)
 698 {
 699         struct lnet_libmd *md = msg->msg_md;
 700
 701         LASSERT(msg->msg_len > 0);
 702         LASSERT(!msg->msg_routing);
 703         LASSERT(md != NULL);
 704         LASSERT(msg->msg_niov == 0);
 705         LASSERT(msg->msg_kiov == NULL);
 706
 707         msg->msg_niov = md->md_niov;
 708         msg->msg_kiov = md->md_kiov;
 709 }
 710
 711 void
 712 lnet_prep_send(struct lnet_msg *msg, int type, struct lnet_process_id target,
 713                unsigned int offset, unsigned int len)
 714 {
 715         msg->msg_type = type;
 716         msg->msg_target = target;
 717         msg->msg_len = len;
 718         msg->msg_offset = offset;
 719
 720         if (len != 0)
 721                 lnet_setpayloadbuffer(msg);
 722
 723         memset (&msg->msg_hdr, 0, sizeof (msg->msg_hdr));
 724         msg->msg_hdr.type           = cpu_to_le32(type);
 725         /* dest_nid will be overwritten by lnet_select_pathway() */
 726         msg->msg_hdr.dest_nid       = cpu_to_le64(target.nid);
 727         msg->msg_hdr.dest_pid       = cpu_to_le32(target.pid);
 728         /* src_nid will be set later */
 729         msg->msg_hdr.src_pid        = cpu_to_le32(the_lnet.ln_pid);
 730         msg->msg_hdr.payload_length = cpu_to_le32(len);
 731 }
 732
 733 void
 734 lnet_ni_send(struct lnet_ni *ni, struct lnet_msg *msg)
 735 {
 736         void *priv = msg->msg_private;
 737         int rc;
 738
 739         LASSERT(!in_interrupt());
 740         LASSERT(nid_is_lo0(&ni->ni_nid) ||
 741                 (msg->msg_txcredit && msg->msg_peertxcredit));
 742
 743         rc = (ni->ni_net->net_lnd->lnd_send)(ni, priv, msg);
 744         if (rc < 0) {
 745                 msg->msg_no_resend = true;
 746                 lnet_finalize(msg, rc);
 747         }
 748 }
 749
 750 static int
 751 lnet_ni_eager_recv(struct lnet_ni *ni, struct lnet_msg *msg)
 752 {
 753         int     rc;
 754
 755         LASSERT(!msg->msg_sending);
 756         LASSERT(msg->msg_receiving);
 757         LASSERT(!msg->msg_rx_ready_delay);
 758         LASSERT(ni->ni_net->net_lnd->lnd_eager_recv != NULL);
 759
 760         msg->msg_rx_ready_delay = 1;
 761         rc = (ni->ni_net->net_lnd->lnd_eager_recv)(ni, msg->msg_private, msg,
 762                                                   &msg->msg_private);
 763         if (rc != 0) {
 764                 CERROR("recv from %s / send to %s aborted: "
 765                        "eager_recv failed %d\n",
 766                        libcfs_nidstr(&msg->msg_rxpeer->lpni_nid),
 767                        libcfs_id2str(msg->msg_target), rc);
 768                 LASSERT(rc < 0); /* required by my callers */
 769         }
 770
 771         return rc;
 772 }
 773
 774 static bool
 775 lnet_is_peer_deadline_passed(struct lnet_peer_ni *lpni, time64_t now)
 776 {
 777         time64_t deadline;
 778
 779         deadline = lpni->lpni_last_alive +
 780                    lpni->lpni_net->net_tunables.lct_peer_timeout;
 781
 782         /*
 783          * assume peer_ni is alive as long as we're within the configured
 784          * peer timeout
 785          */
 786         if (deadline > now)
 787                 return false;
 788
 789         return true;
 790 }
 791
 792 /* NB: returns 1 when alive, 0 when dead, negative when error;
 793  *     may drop the lnet_net_lock */
 794 static int
 795 lnet_peer_alive_locked(struct lnet_ni *ni, struct lnet_peer_ni *lpni,
 796                        struct lnet_msg *msg)
 797 {
 798         time64_t now = ktime_get_seconds();
 799
 800         if (!lnet_peer_aliveness_enabled(lpni))
 801                 return -ENODEV;
 802
 803         /*
 804          * If we're resending a message, let's attempt to send it even if
 805          * the peer is down to fulfill our resend quota on the message
 806          */
 807         if (msg->msg_retry_count > 0)
 808                 return 1;
 809
 810         /* try and send recovery messages irregardless */
 811         if (msg->msg_recovery)
 812                 return 1;
 813
 814         /* always send any responses */
 815         if (lnet_msg_is_response(msg))
 816                 return 1;
 817
 818         if (!lnet_is_peer_deadline_passed(lpni, now))
 819                 return true;
 820
 821         return lnet_is_peer_ni_alive(lpni);
 822 }
 823
 824 /**
 825  * \param msg The message to be sent.
 826  * \param do_send True if lnet_ni_send() should be called in this function.
 827  *        lnet_send() is going to lnet_net_unlock immediately after this, so
 828  *        it sets do_send FALSE and I don't do the unlock/send/lock bit.
 829  *
 830  * \retval LNET_CREDIT_OK If \a msg sent or OK to send.
 831  * \retval LNET_CREDIT_WAIT If \a msg blocked for credit.
 832  * \retval -EHOSTUNREACH If the next hop of the message appears dead.
 833  * \retval -ECANCELED If the MD of the message has been unlinked.
 834  */
 835 static int
 836 lnet_post_send_locked(struct lnet_msg *msg, int do_send)
 837 {
 838         struct lnet_peer_ni     *lp = msg->msg_txpeer;
 839         struct lnet_ni          *ni = msg->msg_txni;
 840         int                     cpt = msg->msg_tx_cpt;
 841         struct lnet_tx_queue    *tq = ni->ni_tx_queues[cpt];
 842
 843         /* non-lnet_send() callers have checked before */
 844         LASSERT(!do_send || msg->msg_tx_delayed);
 845         LASSERT(!msg->msg_receiving);
 846         LASSERT(msg->msg_tx_committed);
 847
 848         /* can't get here if we're sending to the loopback interface */
 849         if (the_lnet.ln_loni)
 850                 LASSERT(!nid_same(&lp->lpni_nid, &the_lnet.ln_loni->ni_nid));
 851
 852         /* NB 'lp' is always the next hop */
 853         if ((msg->msg_target.pid & LNET_PID_USERFLAG) == 0 &&
 854             lnet_peer_alive_locked(ni, lp, msg) == 0) {
 855                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
 856                 the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
 857                         msg->msg_len;
 858                 lnet_net_unlock(cpt);
 859                 if (msg->msg_txpeer)
 860                         lnet_incr_stats(&msg->msg_txpeer->lpni_stats,
 861                                         msg->msg_type,
 862                                         LNET_STATS_TYPE_DROP);
 863                 if (msg->msg_txni)
 864                         lnet_incr_stats(&msg->msg_txni->ni_stats,
 865                                         msg->msg_type,
 866                                         LNET_STATS_TYPE_DROP);
 867
 868                 CNETERR("Dropping message for %s: peer not alive\n",
 869                         libcfs_id2str(msg->msg_target));
 870                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_DROPPED;
 871                 if (do_send)
 872                         lnet_finalize(msg, -EHOSTUNREACH);
 873
 874                 lnet_net_lock(cpt);
 875                 return -EHOSTUNREACH;
 876         }
 877
 878         if (msg->msg_md != NULL &&
 879             (msg->msg_md->md_flags & LNET_MD_FLAG_ABORTED) != 0) {
 880                 lnet_net_unlock(cpt);
 881
 882                 CNETERR("Aborting message for %s: LNetM[DE]Unlink() already "
 883                         "called on the MD/ME.\n",
 884                         libcfs_id2str(msg->msg_target));
 885                 if (do_send) {
 886                         msg->msg_no_resend = true;
 887                         CDEBUG(D_NET, "msg %p to %s canceled and will not be resent\n",
 888                                msg, libcfs_id2str(msg->msg_target));
 889                         lnet_finalize(msg, -ECANCELED);
 890                 }
 891
 892                 lnet_net_lock(cpt);
 893                 return -ECANCELED;
 894         }
 895
 896         if (!msg->msg_peertxcredit) {
 897                 spin_lock(&lp->lpni_lock);
 898                 LASSERT((lp->lpni_txcredits < 0) ==
 899                         !list_empty(&lp->lpni_txq));
 900
 901                 msg->msg_peertxcredit = 1;
 902                 lp->lpni_txqnob += msg->msg_len + sizeof(struct lnet_hdr);
 903                 lp->lpni_txcredits--;
 904
 905                 if (lp->lpni_txcredits < lp->lpni_mintxcredits)
 906                         lp->lpni_mintxcredits = lp->lpni_txcredits;
 907
 908                 if (lp->lpni_txcredits < 0) {
 909                         msg->msg_tx_delayed = 1;
 910                         list_add_tail(&msg->msg_list, &lp->lpni_txq);
 911                         spin_unlock(&lp->lpni_lock);
 912                         return LNET_CREDIT_WAIT;
 913                 }
 914                 spin_unlock(&lp->lpni_lock);
 915         }
 916
 917         if (!msg->msg_txcredit) {
 918                 LASSERT((tq->tq_credits < 0) ==
 919                         !list_empty(&tq->tq_delayed));
 920
 921                 msg->msg_txcredit = 1;
 922                 tq->tq_credits--;
 923                 atomic_dec(&ni->ni_tx_credits);
 924
 925                 if (tq->tq_credits < tq->tq_credits_min)
 926                         tq->tq_credits_min = tq->tq_credits;
 927
 928                 if (tq->tq_credits < 0) {
 929                         msg->msg_tx_delayed = 1;
 930                         list_add_tail(&msg->msg_list, &tq->tq_delayed);
 931                         return LNET_CREDIT_WAIT;
 932                 }
 933         }
 934
 935         if (unlikely(!list_empty(&the_lnet.ln_delay_rules)) &&
 936             lnet_delay_rule_match_locked(&msg->msg_hdr, msg)) {
 937                 msg->msg_tx_delayed = 1;
 938                 return LNET_CREDIT_WAIT;
 939         }
 940
 941         /* unset the tx_delay flag as we're going to send it now */
 942         msg->msg_tx_delayed = 0;
 943
 944         if (do_send) {
 945                 lnet_net_unlock(cpt);
 946                 lnet_ni_send(ni, msg);
 947                 lnet_net_lock(cpt);
 948         }
 949         return LNET_CREDIT_OK;
 950 }
 951
 952
 953 static struct lnet_rtrbufpool *
 954 lnet_msg2bufpool(struct lnet_msg *msg)
 955 {
 956         struct lnet_rtrbufpool  *rbp;
 957         int                     cpt;
 958
 959         LASSERT(msg->msg_rx_committed);
 960
 961         cpt = msg->msg_rx_cpt;
 962         rbp = &the_lnet.ln_rtrpools[cpt][0];
 963
 964         LASSERT(msg->msg_len <= LNET_MTU);
 965         while (msg->msg_len > (unsigned int)rbp->rbp_npages * PAGE_SIZE) {
 966                 rbp++;
 967                 LASSERT(rbp < &the_lnet.ln_rtrpools[cpt][LNET_NRBPOOLS]);
 968         }
 969
 970         return rbp;
 971 }
 972
 973 static int
 974 lnet_post_routed_recv_locked(struct lnet_msg *msg, int do_recv)
 975 {
 976         /* lnet_parse is going to lnet_net_unlock immediately after this, so it
 977          * sets do_recv FALSE and I don't do the unlock/send/lock bit.
 978          * I return LNET_CREDIT_WAIT if msg blocked and LNET_CREDIT_OK if
 979          * received or OK to receive */
 980         struct lnet_peer_ni *lpni = msg->msg_rxpeer;
 981         struct lnet_peer *lp;
 982         struct lnet_rtrbufpool *rbp;
 983         struct lnet_rtrbuf *rb;
 984
 985         LASSERT(msg->msg_kiov == NULL);
 986         LASSERT(msg->msg_niov == 0);
 987         LASSERT(msg->msg_routing);
 988         LASSERT(msg->msg_receiving);
 989         LASSERT(!msg->msg_sending);
 990         LASSERT(lpni->lpni_peer_net);
 991         LASSERT(lpni->lpni_peer_net->lpn_peer);
 992
 993         lp = lpni->lpni_peer_net->lpn_peer;
 994
 995         /* non-lnet_parse callers only receive delayed messages */
 996         LASSERT(!do_recv || msg->msg_rx_delayed);
 997
 998         if (!msg->msg_peerrtrcredit) {
 999                 /* lpni_lock protects the credit manipulation */
1000                 spin_lock(&lpni->lpni_lock);
1001
1002                 msg->msg_peerrtrcredit = 1;
1003                 lpni->lpni_rtrcredits--;
1004                 if (lpni->lpni_rtrcredits < lpni->lpni_minrtrcredits)
1005                         lpni->lpni_minrtrcredits = lpni->lpni_rtrcredits;
1006
1007                 if (lpni->lpni_rtrcredits < 0) {
1008                         spin_unlock(&lpni->lpni_lock);
1009                         /* must have checked eager_recv before here */
1010                         LASSERT(msg->msg_rx_ready_delay);
1011                         msg->msg_rx_delayed = 1;
1012                         /* lp_lock protects the lp_rtrq */
1013                         spin_lock(&lp->lp_lock);
1014                         list_add_tail(&msg->msg_list, &lp->lp_rtrq);
1015                         spin_unlock(&lp->lp_lock);
1016                         return LNET_CREDIT_WAIT;
1017                 }
1018                 spin_unlock(&lpni->lpni_lock);
1019         }
1020
1021         rbp = lnet_msg2bufpool(msg);
1022
1023         if (!msg->msg_rtrcredit) {
1024                 msg->msg_rtrcredit = 1;
1025                 rbp->rbp_credits--;
1026                 if (rbp->rbp_credits < rbp->rbp_mincredits)
1027                         rbp->rbp_mincredits = rbp->rbp_credits;
1028
1029                 if (rbp->rbp_credits < 0) {
1030                         /* must have checked eager_recv before here */
1031                         LASSERT(msg->msg_rx_ready_delay);
1032                         msg->msg_rx_delayed = 1;
1033                         list_add_tail(&msg->msg_list, &rbp->rbp_msgs);
1034                         return LNET_CREDIT_WAIT;
1035                 }
1036         }
1037
1038         LASSERT(!list_empty(&rbp->rbp_bufs));
1039         rb = list_entry(rbp->rbp_bufs.next, struct lnet_rtrbuf, rb_list);
1040         list_del(&rb->rb_list);
1041
1042         msg->msg_niov = rbp->rbp_npages;
1043         msg->msg_kiov = &rb->rb_kiov[0];
1044
1045         /* unset the msg-rx_delayed flag since we're receiving the message */
1046         msg->msg_rx_delayed = 0;
1047
1048         if (do_recv) {
1049                 int cpt = msg->msg_rx_cpt;
1050
1051                 lnet_net_unlock(cpt);
1052                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, msg, 1,
1053                              0, msg->msg_len, msg->msg_len);
1054                 lnet_net_lock(cpt);
1055         }
1056         return LNET_CREDIT_OK;
1057 }
1058
1059 void
1060 lnet_return_tx_credits_locked(struct lnet_msg *msg)
1061 {
1062         struct lnet_peer_ni     *txpeer = msg->msg_txpeer;
1063         struct lnet_ni          *txni = msg->msg_txni;
1064         struct lnet_msg         *msg2;
1065
1066         if (msg->msg_txcredit) {
1067                 struct lnet_ni       *ni = msg->msg_txni;
1068                 struct lnet_tx_queue *tq = ni->ni_tx_queues[msg->msg_tx_cpt];
1069
1070                 /* give back NI txcredits */
1071                 msg->msg_txcredit = 0;
1072
1073                 LASSERT((tq->tq_credits < 0) ==
1074                         !list_empty(&tq->tq_delayed));
1075
1076                 tq->tq_credits++;
1077                 atomic_inc(&ni->ni_tx_credits);
1078                 if (tq->tq_credits <= 0) {
1079                         msg2 = list_entry(tq->tq_delayed.next,
1080                                           struct lnet_msg, msg_list);
1081                         list_del(&msg2->msg_list);
1082
1083                         LASSERT(msg2->msg_txni == ni);
1084                         LASSERT(msg2->msg_tx_delayed);
1085                         LASSERT(msg2->msg_tx_cpt == msg->msg_tx_cpt);
1086
1087                         (void) lnet_post_send_locked(msg2, 1);
1088                 }
1089         }
1090
1091         if (msg->msg_peertxcredit) {
1092                 /* give back peer txcredits */
1093                 msg->msg_peertxcredit = 0;
1094
1095                 spin_lock(&txpeer->lpni_lock);
1096                 LASSERT((txpeer->lpni_txcredits < 0) ==
1097                         !list_empty(&txpeer->lpni_txq));
1098
1099                 txpeer->lpni_txqnob -= msg->msg_len + sizeof(struct lnet_hdr);
1100                 LASSERT(txpeer->lpni_txqnob >= 0);
1101
1102                 txpeer->lpni_txcredits++;
1103                 if (txpeer->lpni_txcredits <= 0) {
1104                         int msg2_cpt;
1105
1106                         msg2 = list_entry(txpeer->lpni_txq.next,
1107                                               struct lnet_msg, msg_list);
1108                         list_del(&msg2->msg_list);
1109                         spin_unlock(&txpeer->lpni_lock);
1110
1111                         LASSERT(msg2->msg_txpeer == txpeer);
1112                         LASSERT(msg2->msg_tx_delayed);
1113
1114                         msg2_cpt = msg2->msg_tx_cpt;
1115
1116                         /*
1117                          * The msg_cpt can be different from the msg2_cpt
1118                          * so we need to make sure we lock the correct cpt
1119                          * for msg2.
1120                          * Once we call lnet_post_send_locked() it is no
1121                          * longer safe to access msg2, since it could've
1122                          * been freed by lnet_finalize(), but we still
1123                          * need to relock the correct cpt, so we cache the
1124                          * msg2_cpt for the purpose of the check that
1125                          * follows the call to lnet_pose_send_locked().
1126                          */
1127                         if (msg2_cpt != msg->msg_tx_cpt) {
1128                                 lnet_net_unlock(msg->msg_tx_cpt);
1129                                 lnet_net_lock(msg2_cpt);
1130                         }
1131                         (void) lnet_post_send_locked(msg2, 1);
1132                         if (msg2_cpt != msg->msg_tx_cpt) {
1133                                 lnet_net_unlock(msg2_cpt);
1134                                 lnet_net_lock(msg->msg_tx_cpt);
1135                         }
1136                 } else {
1137                         spin_unlock(&txpeer->lpni_lock);
1138                 }
1139         }
1140
1141         if (txni != NULL) {
1142                 msg->msg_txni = NULL;
1143                 lnet_ni_decref_locked(txni, msg->msg_tx_cpt);
1144         }
1145
1146         if (txpeer != NULL) {
1147                 msg->msg_txpeer = NULL;
1148                 lnet_peer_ni_decref_locked(txpeer);
1149         }
1150 }
1151
1152 void
1153 lnet_schedule_blocked_locked(struct lnet_rtrbufpool *rbp)
1154 {
1155         struct lnet_msg *msg;
1156
1157         if (list_empty(&rbp->rbp_msgs))
1158                 return;
1159         msg = list_entry(rbp->rbp_msgs.next,
1160                          struct lnet_msg, msg_list);
1161         list_del(&msg->msg_list);
1162
1163         (void)lnet_post_routed_recv_locked(msg, 1);
1164 }
1165
1166 void
1167 lnet_drop_routed_msgs_locked(struct list_head *list, int cpt)
1168 {
1169         struct lnet_msg *msg;
1170         struct lnet_msg *tmp;
1171
1172         lnet_net_unlock(cpt);
1173
1174         list_for_each_entry_safe(msg, tmp, list, msg_list) {
1175                 lnet_ni_recv(msg->msg_rxni, msg->msg_private, NULL,
1176                              0, 0, 0, msg->msg_hdr.payload_length);
1177                 list_del_init(&msg->msg_list);
1178                 msg->msg_no_resend = true;
1179                 msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
1180                 lnet_finalize(msg, -ECANCELED);
1181         }
1182
1183         lnet_net_lock(cpt);
1184 }
1185
1186 void
1187 lnet_return_rx_credits_locked(struct lnet_msg *msg)
1188 {
1189         struct lnet_peer_ni *rxpeerni = msg->msg_rxpeer;
1190         struct lnet_peer *lp;
1191         struct lnet_ni *rxni = msg->msg_rxni;
1192         struct lnet_msg *msg2;
1193
1194         if (msg->msg_rtrcredit) {
1195                 /* give back global router credits */
1196                 struct lnet_rtrbuf *rb;
1197                 struct lnet_rtrbufpool *rbp;
1198
1199                 /* NB If a msg ever blocks for a buffer in rbp_msgs, it stays
1200                  * there until it gets one allocated, or aborts the wait
1201                  * itself */
1202                 LASSERT(msg->msg_kiov != NULL);
1203
1204                 rb = list_entry(msg->msg_kiov, struct lnet_rtrbuf, rb_kiov[0]);
1205                 rbp = rb->rb_pool;
1206
1207                 msg->msg_kiov = NULL;
1208                 msg->msg_rtrcredit = 0;
1209
1210                 LASSERT(rbp == lnet_msg2bufpool(msg));
1211
1212                 LASSERT((rbp->rbp_credits > 0) ==
1213                         !list_empty(&rbp->rbp_bufs));
1214
1215                 /* If routing is now turned off, we just drop this buffer and
1216                  * don't bother trying to return credits.  */
1217                 if (!the_lnet.ln_routing) {
1218                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1219                         goto routing_off;
1220                 }
1221
1222                 /* It is possible that a user has lowered the desired number of
1223                  * buffers in this pool.  Make sure we never put back
1224                  * more buffers than the stated number. */
1225                 if (unlikely(rbp->rbp_credits >= rbp->rbp_req_nbuffers)) {
1226                         /* Discard this buffer so we don't have too
1227                          * many. */
1228                         lnet_destroy_rtrbuf(rb, rbp->rbp_npages);
1229                         rbp->rbp_nbuffers--;
1230                 } else {
1231                         list_add(&rb->rb_list, &rbp->rbp_bufs);
1232                         rbp->rbp_credits++;
1233                         if (rbp->rbp_credits <= 0)
1234                                 lnet_schedule_blocked_locked(rbp);
1235                 }
1236         }
1237
1238 routing_off:
1239         if (msg->msg_peerrtrcredit) {
1240                 LASSERT(rxpeerni);
1241                 LASSERT(rxpeerni->lpni_peer_net);
1242                 LASSERT(rxpeerni->lpni_peer_net->lpn_peer);
1243
1244                 /* give back peer router credits */
1245                 msg->msg_peerrtrcredit = 0;
1246
1247                 spin_lock(&rxpeerni->lpni_lock);
1248                 rxpeerni->lpni_rtrcredits++;
1249                 spin_unlock(&rxpeerni->lpni_lock);
1250
1251                 lp = rxpeerni->lpni_peer_net->lpn_peer;
1252                 spin_lock(&lp->lp_lock);
1253
1254                 /* drop all messages which are queued to be routed on that
1255                  * peer. */
1256                 if (!the_lnet.ln_routing) {
1257                         LIST_HEAD(drop);
1258                         list_splice_init(&lp->lp_rtrq, &drop);
1259                         spin_unlock(&lp->lp_lock);
1260                         lnet_drop_routed_msgs_locked(&drop, msg->msg_rx_cpt);
1261                 } else if (!list_empty(&lp->lp_rtrq)) {
1262                         int msg2_cpt;
1263
1264                         msg2 = list_entry(lp->lp_rtrq.next,
1265                                           struct lnet_msg, msg_list);
1266                         list_del(&msg2->msg_list);
1267                         msg2_cpt = msg2->msg_rx_cpt;
1268                         spin_unlock(&lp->lp_lock);
1269                         /*
1270                          * messages on the lp_rtrq can be from any NID in
1271                          * the peer, which means they might have different
1272                          * cpts. We need to make sure we lock the right
1273                          * one.
1274                          */
1275                         if (msg2_cpt != msg->msg_rx_cpt) {
1276                                 lnet_net_unlock(msg->msg_rx_cpt);
1277                                 lnet_net_lock(msg2_cpt);
1278                         }
1279                         (void) lnet_post_routed_recv_locked(msg2, 1);
1280                         if (msg2_cpt != msg->msg_rx_cpt) {
1281                                 lnet_net_unlock(msg2_cpt);
1282                                 lnet_net_lock(msg->msg_rx_cpt);
1283                         }
1284                 } else {
1285                         spin_unlock(&lp->lp_lock);
1286                 }
1287         }
1288         if (rxni != NULL) {
1289                 msg->msg_rxni = NULL;
1290                 lnet_ni_decref_locked(rxni, msg->msg_rx_cpt);
1291         }
1292         if (rxpeerni != NULL) {
1293                 msg->msg_rxpeer = NULL;
1294                 lnet_peer_ni_decref_locked(rxpeerni);
1295         }
1296 }
1297
1298 static struct lnet_peer_ni *
1299 lnet_select_peer_ni(struct lnet_ni *best_ni, lnet_nid_t dst_nid,
1300                     struct lnet_peer *peer,
1301                     struct lnet_peer_ni *best_lpni,
1302                     struct lnet_peer_net *peer_net)
1303 {
1304         /*
1305          * Look at the peer NIs for the destination peer that connect
1306          * to the chosen net. If a peer_ni is preferred when using the
1307          * best_ni to communicate, we use that one. If there is no
1308          * preferred peer_ni, or there are multiple preferred peer_ni,
1309          * the available transmit credits are used. If the transmit
1310          * credits are equal, we round-robin over the peer_ni.
1311          */
1312         struct lnet_peer_ni *lpni = NULL;
1313         int best_lpni_credits = (best_lpni) ? best_lpni->lpni_txcredits :
1314                 INT_MIN;
1315         int best_lpni_healthv = (best_lpni) ?
1316                 atomic_read(&best_lpni->lpni_healthv) : 0;
1317         bool best_lpni_is_preferred = false;
1318         bool lpni_is_preferred;
1319         int lpni_healthv;
1320         __u32 lpni_sel_prio;
1321         __u32 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1322
1323         while ((lpni = lnet_get_next_peer_ni_locked(peer, peer_net, lpni))) {
1324                 /*
1325                  * if the best_ni we've chosen aleady has this lpni
1326                  * preferred, then let's use it
1327                  */
1328                 if (best_ni) {
1329                         lpni_is_preferred = lnet_peer_is_pref_nid_locked(
1330                                 lpni, &best_ni->ni_nid);
1331                         CDEBUG(D_NET, "%s lpni_is_preferred = %d\n",
1332                                libcfs_nidstr(&best_ni->ni_nid),
1333                                lpni_is_preferred);
1334                 } else {
1335                         lpni_is_preferred = false;
1336                 }
1337
1338                 lpni_healthv = atomic_read(&lpni->lpni_healthv);
1339                 lpni_sel_prio = lpni->lpni_sel_priority;
1340
1341                 if (best_lpni)
1342                         CDEBUG(D_NET, "n:[%s, %s] h:[%d, %d] p:[%d, %d] c:[%d, %d] s:[%d, %d]\n",
1343                                 libcfs_nidstr(&lpni->lpni_nid),
1344                                 libcfs_nidstr(&best_lpni->lpni_nid),
1345                                 lpni_healthv, best_lpni_healthv,
1346                                 lpni_sel_prio, best_sel_prio,
1347                                 lpni->lpni_txcredits, best_lpni_credits,
1348                                 lpni->lpni_seq, best_lpni->lpni_seq);
1349                 else
1350                         goto select_lpni;
1351
1352                 /* pick the healthiest peer ni */
1353                 if (lpni_healthv < best_lpni_healthv)
1354                         continue;
1355                 else if (lpni_healthv > best_lpni_healthv) {
1356                         if (best_lpni_is_preferred)
1357                                 best_lpni_is_preferred = false;
1358                         goto select_lpni;
1359                 }
1360
1361                 if (lpni_sel_prio > best_sel_prio)
1362                         continue;
1363                 else if (lpni_sel_prio < best_sel_prio) {
1364                         if (best_lpni_is_preferred)
1365                                 best_lpni_is_preferred = false;
1366                         goto select_lpni;
1367                 }
1368
1369                 /* if this is a preferred peer use it */
1370                 if (!best_lpni_is_preferred && lpni_is_preferred) {
1371                         best_lpni_is_preferred = true;
1372                         goto select_lpni;
1373                 } else if (best_lpni_is_preferred && !lpni_is_preferred) {
1374                         /* this is not the preferred peer so let's ignore
1375                          * it.
1376                          */
1377                         continue;
1378                 }
1379
1380                 if (lpni->lpni_txcredits < best_lpni_credits)
1381                         /* We already have a peer that has more credits
1382                          * available than this one. No need to consider
1383                          * this peer further.
1384                          */
1385                         continue;
1386                 else if (lpni->lpni_txcredits > best_lpni_credits)
1387                         goto select_lpni;
1388
1389                 /* The best peer found so far and the current peer
1390                  * have the same number of available credits let's
1391                  * make sure to select between them using Round Robin
1392                  */
1393                 if (best_lpni && (best_lpni->lpni_seq <= lpni->lpni_seq))
1394                         continue;
1395 select_lpni:
1396                 best_lpni_is_preferred = lpni_is_preferred;
1397                 best_lpni_healthv = lpni_healthv;
1398                 best_sel_prio = lpni_sel_prio;
1399                 best_lpni = lpni;
1400                 best_lpni_credits = lpni->lpni_txcredits;
1401         }
1402
1403         /* if we still can't find a peer ni then we can't reach it */
1404         if (!best_lpni) {
1405                 __u32 net_id = (peer_net) ? peer_net->lpn_net_id :
1406                         LNET_NIDNET(dst_nid);
1407                 CDEBUG(D_NET, "no peer_ni found on peer net %s\n",
1408                                 libcfs_net2str(net_id));
1409                 return NULL;
1410         }
1411
1412         CDEBUG(D_NET, "sd_best_lpni = %s\n",
1413                libcfs_nidstr(&best_lpni->lpni_nid));
1414
1415         return best_lpni;
1416 }
1417
1418 /*
1419  * Prerequisite: the best_ni should already be set in the sd
1420  * Find the best lpni.
1421  * If the net id is provided then restrict lpni selection on
1422  * that particular net.
1423  * Otherwise find any reachable lpni. When dealing with an MR
1424  * gateway and it has multiple lpnis which we can use
1425  * we want to select the best one from the list of reachable
1426  * ones.
1427  */
1428 static inline struct lnet_peer_ni *
1429 lnet_find_best_lpni(struct lnet_ni *lni, lnet_nid_t dst_nid,
1430                     struct lnet_peer *peer, __u32 net_id)
1431 {
1432         struct lnet_peer_net *peer_net;
1433
1434         /* find the best_lpni on any local network */
1435         if (net_id == LNET_NET_ANY) {
1436                 struct lnet_peer_ni *best_lpni = NULL;
1437                 struct lnet_peer_net *lpn;
1438                 list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
1439                         /* no net specified find any reachable peer ni */
1440                         if (!lnet_islocalnet_locked(lpn->lpn_net_id))
1441                                 continue;
1442                         best_lpni = lnet_select_peer_ni(lni, dst_nid, peer,
1443                                                         best_lpni, lpn);
1444                 }
1445
1446                 return best_lpni;
1447         }
1448         /* restrict on the specified net */
1449         peer_net = lnet_peer_get_net_locked(peer, net_id);
1450         if (peer_net)
1451                 return lnet_select_peer_ni(lni, dst_nid, peer, NULL, peer_net);
1452
1453         return NULL;
1454 }
1455
1456 static int
1457 lnet_compare_gw_lpnis(struct lnet_peer_ni *lpni1, struct lnet_peer_ni *lpni2)
1458 {
1459         if (lpni1->lpni_txqnob < lpni2->lpni_txqnob)
1460                 return 1;
1461
1462         if (lpni1->lpni_txqnob > lpni2->lpni_txqnob)
1463                 return -1;
1464
1465         if (lpni1->lpni_txcredits > lpni2->lpni_txcredits)
1466                 return 1;
1467
1468         if (lpni1->lpni_txcredits < lpni2->lpni_txcredits)
1469                 return -1;
1470
1471         return 0;
1472 }
1473
1474 /* Compare route priorities and hop counts */
1475 static int
1476 lnet_compare_routes(struct lnet_route *r1, struct lnet_route *r2)
1477 {
1478         int r1_hops = (r1->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r1->lr_hops;
1479         int r2_hops = (r2->lr_hops == LNET_UNDEFINED_HOPS) ? 1 : r2->lr_hops;
1480
1481         if (r1->lr_priority < r2->lr_priority)
1482                 return 1;
1483
1484         if (r1->lr_priority > r2->lr_priority)
1485                 return -1;
1486
1487         if (r1_hops < r2_hops)
1488                 return 1;
1489
1490         if (r1_hops > r2_hops)
1491                 return -1;
1492
1493         return 0;
1494 }
1495
1496 static struct lnet_route *
1497 lnet_find_route_locked(struct lnet_remotenet *rnet, __u32 src_net,
1498                        struct lnet_peer_ni *remote_lpni,
1499                        struct lnet_route **prev_route,
1500                        struct lnet_peer_ni **gwni)
1501 {
1502         struct lnet_peer_ni *lpni, *best_gw_ni = NULL;
1503         struct lnet_route *best_route;
1504         struct lnet_route *last_route;
1505         struct lnet_route *route;
1506         int rc;
1507         bool best_rte_is_preferred = false;
1508         struct lnet_nid *gw_pnid;
1509
1510         CDEBUG(D_NET, "Looking up a route to %s, from %s\n",
1511                libcfs_net2str(rnet->lrn_net), libcfs_net2str(src_net));
1512
1513         best_route = last_route = NULL;
1514         list_for_each_entry(route, &rnet->lrn_routes, lr_list) {
1515                 if (!lnet_is_route_alive(route))
1516                         continue;
1517                 gw_pnid = &route->lr_gateway->lp_primary_nid;
1518
1519                 /* no protection on below fields, but it's harmless */
1520                 if (last_route && (last_route->lr_seq - route->lr_seq < 0))
1521                         last_route = route;
1522
1523                 /* if the best route found is in the preferred list then
1524                  * tag it as preferred and use it later on. But if we
1525                  * didn't find any routes which are on the preferred list
1526                  * then just use the best route possible.
1527                  */
1528                 rc = lnet_peer_is_pref_rtr_locked(remote_lpni, gw_pnid);
1529
1530                 if (!best_route || (rc && !best_rte_is_preferred)) {
1531                         /* Restrict the selection of the router NI on the
1532                          * src_net provided. If the src_net is LNET_NID_ANY,
1533                          * then select the best interface available.
1534                          */
1535                         lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1536                                                    route->lr_gateway,
1537                                                    src_net);
1538                         if (!lpni) {
1539                                 CDEBUG(D_NET,
1540                                        "Gateway %s does not have a peer NI on net %s\n",
1541                                        libcfs_nidstr(gw_pnid),
1542                                        libcfs_net2str(src_net));
1543                                 continue;
1544                         }
1545                 }
1546
1547                 if (rc && !best_rte_is_preferred) {
1548                         /* This is the first preferred route we found,
1549                          * so it beats any route found previously
1550                          */
1551                         best_route = route;
1552                         if (!last_route)
1553                                 last_route = route;
1554                         best_gw_ni = lpni;
1555                         best_rte_is_preferred = true;
1556                         CDEBUG(D_NET, "preferred gw = %s\n",
1557                                libcfs_nidstr(gw_pnid));
1558                         continue;
1559                 } else if ((!rc) && best_rte_is_preferred)
1560                         /* The best route we found so far is in the preferred
1561                          * list, so it beats any non-preferred route
1562                          */
1563                         continue;
1564
1565                 if (!best_route) {
1566                         best_route = last_route = route;
1567                         best_gw_ni = lpni;
1568                         continue;
1569                 }
1570
1571                 rc = lnet_compare_routes(route, best_route);
1572                 if (rc == -1)
1573                         continue;
1574
1575                 /* Restrict the selection of the router NI on the
1576                  * src_net provided. If the src_net is LNET_NID_ANY,
1577                  * then select the best interface available.
1578                  */
1579                 lpni = lnet_find_best_lpni(NULL, LNET_NID_ANY,
1580                                            route->lr_gateway,
1581                                            src_net);
1582                 if (!lpni) {
1583                         CDEBUG(D_NET,
1584                                "Gateway %s does not have a peer NI on net %s\n",
1585                                libcfs_nidstr(gw_pnid),
1586                                libcfs_net2str(src_net));
1587                         continue;
1588                 }
1589
1590                 if (rc == 1) {
1591                         best_route = route;
1592                         best_gw_ni = lpni;
1593                         continue;
1594                 }
1595
1596                 rc = lnet_compare_gw_lpnis(lpni, best_gw_ni);
1597                 if (rc == -1)
1598                         continue;
1599
1600                 if (rc == 1 || route->lr_seq <= best_route->lr_seq) {
1601                         best_route = route;
1602                         best_gw_ni = lpni;
1603                         continue;
1604                 }
1605         }
1606
1607         *prev_route = last_route;
1608         *gwni = best_gw_ni;
1609
1610         return best_route;
1611 }
1612
1613 static inline unsigned int
1614 lnet_dev_prio_of_md(struct lnet_ni *ni, unsigned int dev_idx)
1615 {
1616         if (dev_idx == UINT_MAX)
1617                 return UINT_MAX;
1618
1619         if (!ni || !ni->ni_net || !ni->ni_net->net_lnd ||
1620             !ni->ni_net->net_lnd->lnd_get_dev_prio)
1621                 return UINT_MAX;
1622
1623         return ni->ni_net->net_lnd->lnd_get_dev_prio(ni, dev_idx);
1624 }
1625
1626 static struct lnet_ni *
1627 lnet_get_best_ni(struct lnet_net *local_net, struct lnet_ni *best_ni,
1628                  struct lnet_peer *peer, struct lnet_peer_net *peer_net,
1629                  struct lnet_msg *msg, int md_cpt)
1630 {
1631         struct lnet_libmd *md = msg->msg_md;
1632         unsigned int offset = msg->msg_offset;
1633         unsigned int shortest_distance;
1634         struct lnet_ni *ni = NULL;
1635         int best_credits;
1636         int best_healthv;
1637         __u32 best_sel_prio;
1638         unsigned int best_dev_prio;
1639         unsigned int dev_idx = UINT_MAX;
1640         struct page *page = lnet_get_first_page(md, offset);
1641         msg->msg_rdma_force = lnet_is_rdma_only_page(page);
1642
1643         if (msg->msg_rdma_force)
1644                 dev_idx = lnet_get_dev_idx(page);
1645
1646         /*
1647          * If there is no peer_ni that we can send to on this network,
1648          * then there is no point in looking for a new best_ni here.
1649         */
1650         if (!lnet_get_next_peer_ni_locked(peer, peer_net, NULL))
1651                 return best_ni;
1652
1653         if (best_ni == NULL) {
1654                 best_sel_prio = LNET_MAX_SELECTION_PRIORITY;
1655                 shortest_distance = UINT_MAX;
1656                 best_dev_prio = UINT_MAX;
1657                 best_credits = INT_MIN;
1658                 best_healthv = 0;
1659         } else {
1660                 best_dev_prio = lnet_dev_prio_of_md(best_ni, dev_idx);
1661                 shortest_distance = cfs_cpt_distance(lnet_cpt_table(), md_cpt,
1662                                                      best_ni->ni_dev_cpt);
1663                 best_credits = atomic_read(&best_ni->ni_tx_credits);
1664                 best_healthv = atomic_read(&best_ni->ni_healthv);
1665                 best_sel_prio = best_ni->ni_sel_priority;
1666         }
1667
1668         while ((ni = lnet_get_next_ni_locked(local_net, ni))) {
1669                 unsigned int distance;
1670                 int ni_credits;
1671                 int ni_healthv;
1672                 int ni_fatal;
1673                 __u32 ni_sel_prio;
1674                 unsigned int ni_dev_prio;
1675
1676                 ni_credits = atomic_read(&ni->ni_tx_credits);
1677                 ni_healthv = atomic_read(&ni->ni_healthv);
1678                 ni_fatal = atomic_read(&ni->ni_fatal_error_on);
1679                 ni_sel_prio = ni->ni_sel_priority;
1680
1681                 /*
1682                  * calculate the distance from the CPT on which
1683                  * the message memory is allocated to the CPT of
1684                  * the NI's physical device
1685                  */
1686                 distance = cfs_cpt_distance(lnet_cpt_table(),
1687                                             md_cpt,
1688                                             ni->ni_dev_cpt);
1689
1690                 ni_dev_prio = lnet_dev_prio_of_md(ni, dev_idx);
1691
1692                 /*
1693                  * All distances smaller than the NUMA range
1694                  * are treated equally.
1695                  */
1696                 if (distance < lnet_numa_range)
1697                         distance = lnet_numa_range;
1698
1699                 /*
1700                  * Select on health, selection policy, direct dma prio,
1701                  * shorter distance, available credits, then round-robin.
1702                  */
1703                 if (ni_fatal)
1704                         continue;
1705
1706                 if (best_ni)
1707                         CDEBUG(D_NET, "compare ni %s [c:%d, d:%d, s:%d, p:%u, g:%u] with best_ni %s [c:%d, d:%d, s:%d, p:%u, g:%u]\n",
1708                                libcfs_nidstr(&ni->ni_nid), ni_credits, distance,
1709                                ni->ni_seq, ni_sel_prio, ni_dev_prio,
1710                                (best_ni) ? libcfs_nidstr(&best_ni->ni_nid)
1711                                : "not selected", best_credits, shortest_distance,
1712                                (best_ni) ? best_ni->ni_seq : 0,
1713                                best_sel_prio, best_dev_prio);
1714                 else
1715                         goto select_ni;
1716
1717                 if (ni_healthv < best_healthv)
1718                         continue;
1719                 else if (ni_healthv > best_healthv)
1720                         goto select_ni;
1721
1722                 if (ni_sel_prio > best_sel_prio)
1723                         continue;
1724                 else if (ni_sel_prio < best_sel_prio)
1725                         goto select_ni;
1726
1727                 if (ni_dev_prio > best_dev_prio)
1728                         continue;
1729                 else if (ni_dev_prio < best_dev_prio)
1730                         goto select_ni;
1731
1732                 if (distance > shortest_distance)
1733                         continue;
1734                 else if (distance < shortest_distance)
1735                         goto select_ni;
1736
1737                 if (ni_credits < best_credits)
1738                         continue;
1739                 else if (ni_credits > best_credits)
1740                         goto select_ni;
1741
1742                 if (best_ni && best_ni->ni_seq <= ni->ni_seq)
1743                         continue;
1744
1745 select_ni:
1746                 best_sel_prio = ni_sel_prio;
1747                 best_dev_prio = ni_dev_prio;
1748                 shortest_distance = distance;
1749                 best_healthv = ni_healthv;
1750                 best_ni = ni;
1751                 best_credits = ni_credits;
1752         }
1753
1754         CDEBUG(D_NET, "selected best_ni %s\n",
1755                (best_ni) ? libcfs_nidstr(&best_ni->ni_nid) : "no selection");
1756
1757         return best_ni;
1758 }
1759
1760 /*
1761  * Traffic to the LNET_RESERVED_PORTAL may not trigger peer discovery,
1762  * because such traffic is required to perform discovery. We therefore
1763  * exclude all GET and PUT on that portal. We also exclude all ACK and
1764  * REPLY traffic, but that is because the portal is not tracked in the
1765  * message structure for these message types. We could restrict this
1766  * further by also checking for LNET_PROTO_PING_MATCHBITS.
1767  */
1768 static bool
1769 lnet_msg_discovery(struct lnet_msg *msg)
1770 {
1771         if (msg->msg_type == LNET_MSG_PUT) {
1772                 if (msg->msg_hdr.msg.put.ptl_index != LNET_RESERVED_PORTAL)
1773                         return true;
1774         } else if (msg->msg_type == LNET_MSG_GET) {
1775                 if (msg->msg_hdr.msg.get.ptl_index != LNET_RESERVED_PORTAL)
1776                         return true;
1777         }
1778         return false;
1779 }
1780
1781 #define SRC_SPEC        0x0001
1782 #define SRC_ANY         0x0002
1783 #define LOCAL_DST       0x0004
1784 #define REMOTE_DST      0x0008
1785 #define MR_DST          0x0010
1786 #define NMR_DST         0x0020
1787 #define SND_RESP        0x0040
1788
1789 /* The following to defines are used for return codes */
1790 #define REPEAT_SEND     0x1000
1791 #define PASS_THROUGH    0x2000
1792
1793 /* The different cases lnet_select pathway needs to handle */
1794 #define SRC_SPEC_LOCAL_MR_DST   (SRC_SPEC | LOCAL_DST | MR_DST)
1795 #define SRC_SPEC_ROUTER_MR_DST  (SRC_SPEC | REMOTE_DST | MR_DST)
1796 #define SRC_SPEC_LOCAL_NMR_DST  (SRC_SPEC | LOCAL_DST | NMR_DST)
1797 #define SRC_SPEC_ROUTER_NMR_DST (SRC_SPEC | REMOTE_DST | NMR_DST)
1798 #define SRC_ANY_LOCAL_MR_DST    (SRC_ANY | LOCAL_DST | MR_DST)
1799 #define SRC_ANY_ROUTER_MR_DST   (SRC_ANY | REMOTE_DST | MR_DST)
1800 #define SRC_ANY_LOCAL_NMR_DST   (SRC_ANY | LOCAL_DST | NMR_DST)
1801 #define SRC_ANY_ROUTER_NMR_DST  (SRC_ANY | REMOTE_DST | NMR_DST)
1802
1803 static int
1804 lnet_handle_lo_send(struct lnet_send_data *sd)
1805 {
1806         struct lnet_msg *msg = sd->sd_msg;
1807         int cpt = sd->sd_cpt;
1808
1809         if (the_lnet.ln_state != LNET_STATE_RUNNING)
1810                 return -ESHUTDOWN;
1811
1812         /* No send credit hassles with LOLND */
1813         lnet_ni_addref_locked(the_lnet.ln_loni, cpt);
1814         msg->msg_hdr.dest_nid =
1815                 cpu_to_le64(lnet_nid_to_nid4(&the_lnet.ln_loni->ni_nid));
1816         if (!msg->msg_routing)
1817                 msg->msg_hdr.src_nid =
1818                         cpu_to_le64(lnet_nid_to_nid4(&the_lnet.ln_loni->ni_nid));
1819         msg->msg_target.nid = lnet_nid_to_nid4(&the_lnet.ln_loni->ni_nid);
1820         lnet_msg_commit(msg, cpt);
1821         msg->msg_txni = the_lnet.ln_loni;
1822
1823         return LNET_CREDIT_OK;
1824 }
1825
1826 static int
1827 lnet_handle_send(struct lnet_send_data *sd)
1828 {
1829         struct lnet_ni *best_ni = sd->sd_best_ni;
1830         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
1831         struct lnet_peer_ni *final_dst_lpni = sd->sd_final_dst_lpni;
1832         struct lnet_msg *msg = sd->sd_msg;
1833         int cpt2;
1834         __u32 send_case = sd->sd_send_case;
1835         int rc;
1836         __u32 routing = send_case & REMOTE_DST;
1837          struct lnet_rsp_tracker *rspt;
1838
1839         /* Increment sequence number of the selected peer, peer net,
1840          * local ni and local net so that we pick the next ones
1841          * in Round Robin.
1842          */
1843         best_lpni->lpni_peer_net->lpn_seq++;
1844         best_lpni->lpni_seq = best_lpni->lpni_peer_net->lpn_seq;
1845         best_ni->ni_net->net_seq++;
1846         best_ni->ni_seq = best_ni->ni_net->net_seq;
1847
1848         CDEBUG(D_NET, "%s NI seq info: [%d:%d:%d:%u] %s LPNI seq info [%d:%d:%d:%u]\n",
1849                libcfs_nidstr(&best_ni->ni_nid),
1850                best_ni->ni_seq, best_ni->ni_net->net_seq,
1851                atomic_read(&best_ni->ni_tx_credits),
1852                best_ni->ni_sel_priority,
1853                libcfs_nidstr(&best_lpni->lpni_nid),
1854                best_lpni->lpni_seq, best_lpni->lpni_peer_net->lpn_seq,
1855                best_lpni->lpni_txcredits,
1856                best_lpni->lpni_sel_priority);
1857
1858         /*
1859          * grab a reference on the peer_ni so it sticks around even if
1860          * we need to drop and relock the lnet_net_lock below.
1861          */
1862         lnet_peer_ni_addref_locked(best_lpni);
1863
1864         /*
1865          * Use lnet_cpt_of_nid() to determine the CPT used to commit the
1866          * message. This ensures that we get a CPT that is correct for
1867          * the NI when the NI has been restricted to a subset of all CPTs.
1868          * If the selected CPT differs from the one currently locked, we
1869          * must unlock and relock the lnet_net_lock(), and then check whether
1870          * the configuration has changed. We don't have a hold on the best_ni
1871          * yet, and it may have vanished.
1872          */
1873         cpt2 = lnet_cpt_of_nid_locked(&best_lpni->lpni_nid, best_ni);
1874         if (sd->sd_cpt != cpt2) {
1875                 __u32 seq = lnet_get_dlc_seq_locked();
1876                 lnet_net_unlock(sd->sd_cpt);
1877                 sd->sd_cpt = cpt2;
1878                 lnet_net_lock(sd->sd_cpt);
1879                 if (seq != lnet_get_dlc_seq_locked()) {
1880                         lnet_peer_ni_decref_locked(best_lpni);
1881                         return REPEAT_SEND;
1882                 }
1883         }
1884
1885         /*
1886          * store the best_lpni in the message right away to avoid having
1887          * to do the same operation under different conditions
1888          */
1889         msg->msg_txpeer = best_lpni;
1890         msg->msg_txni = best_ni;
1891
1892         /*
1893          * grab a reference for the best_ni since now it's in use in this
1894          * send. The reference will be dropped in lnet_finalize()
1895          */
1896         lnet_ni_addref_locked(msg->msg_txni, sd->sd_cpt);
1897
1898         /*
1899          * Always set the target.nid to the best peer picked. Either the
1900          * NID will be one of the peer NIDs selected, or the same NID as
1901          * what was originally set in the target or it will be the NID of
1902          * a router if this message should be routed
1903          */
1904         /* FIXME handle large-addr nids */
1905         msg->msg_target.nid = lnet_nid_to_nid4(&msg->msg_txpeer->lpni_nid);
1906
1907         /*
1908          * lnet_msg_commit assigns the correct cpt to the message, which
1909          * is used to decrement the correct refcount on the ni when it's
1910          * time to return the credits
1911          */
1912         lnet_msg_commit(msg, sd->sd_cpt);
1913
1914         /*
1915          * If we are routing the message then we keep the src_nid that was
1916          * set by the originator. If we are not routing then we are the
1917          * originator and set it here.
1918          */
1919         if (!msg->msg_routing)
1920                 msg->msg_hdr.src_nid =
1921                         cpu_to_le64(lnet_nid_to_nid4(&msg->msg_txni->ni_nid));
1922
1923         if (routing) {
1924                 msg->msg_target_is_router = 1;
1925                 msg->msg_target.pid = LNET_PID_LUSTRE;
1926                 /*
1927                  * since we're routing we want to ensure that the
1928                  * msg_hdr.dest_nid is set to the final destination. When
1929                  * the router receives this message it knows how to route
1930                  * it.
1931                  *
1932                  * final_dst_lpni is set at the beginning of the
1933                  * lnet_select_pathway() function and is never changed.
1934                  * It's safe to use it here.
1935                  */
1936                 /* FIXME handle large-addr nid */
1937                 msg->msg_hdr.dest_nid =
1938                         cpu_to_le64(lnet_nid_to_nid4(&final_dst_lpni->lpni_nid));
1939         } else {
1940                 /*
1941                  * if we're not routing set the dest_nid to the best peer
1942                  * ni NID that we picked earlier in the algorithm.
1943                  */
1944                 msg->msg_hdr.dest_nid =
1945                         cpu_to_le64(lnet_nid_to_nid4(&msg->msg_txpeer->lpni_nid));
1946         }
1947
1948         /*
1949          * if we have response tracker block update it with the next hop
1950          * nid
1951          */
1952         if (msg->msg_md) {
1953                 rspt = msg->msg_md->md_rspt_ptr;
1954                 if (rspt) {
1955                         rspt->rspt_next_hop_nid =
1956                                 msg->msg_txpeer->lpni_nid;
1957                         CDEBUG(D_NET, "rspt_next_hop_nid = %s\n",
1958                                libcfs_nidstr(&rspt->rspt_next_hop_nid));
1959                 }
1960         }
1961
1962         rc = lnet_post_send_locked(msg, 0);
1963
1964         if (!rc)
1965                 CDEBUG(D_NET, "TRACE: %s(%s:%s) -> %s(%s:%s) %s : %s try# %d\n",
1966                        libcfs_nid2str(msg->msg_hdr.src_nid),
1967                        libcfs_nidstr(&msg->msg_txni->ni_nid),
1968                        libcfs_nid2str(sd->sd_src_nid),
1969                        libcfs_nid2str(msg->msg_hdr.dest_nid),
1970                        libcfs_nid2str(sd->sd_dst_nid),
1971                        libcfs_nidstr(&msg->msg_txpeer->lpni_nid),
1972                        libcfs_nid2str(sd->sd_rtr_nid),
1973                        lnet_msgtyp2str(msg->msg_type), msg->msg_retry_count);
1974
1975         return rc;
1976 }
1977
1978 static inline void
1979 lnet_set_non_mr_pref_nid(struct lnet_peer_ni *lpni, struct lnet_ni *lni,
1980                          struct lnet_msg *msg)
1981 {
1982         if (!lnet_peer_is_multi_rail(lpni->lpni_peer_net->lpn_peer) &&
1983             !lnet_msg_is_response(msg) && lpni->lpni_pref_nnids == 0) {
1984                 CDEBUG(D_NET, "Setting preferred local NID %s on NMR peer %s\n",
1985                        libcfs_nidstr(&lni->ni_nid),
1986                        libcfs_nidstr(&lpni->lpni_nid));
1987                 lnet_peer_ni_set_non_mr_pref_nid(lpni, &lni->ni_nid);
1988         }
1989 }
1990
1991 /*
1992  * Source Specified
1993  * Local Destination
1994  * non-mr peer
1995  *
1996  * use the source and destination NIDs as the pathway
1997  */
1998 static int
1999 lnet_handle_spec_local_nmr_dst(struct lnet_send_data *sd)
2000 {
2001         /* the destination lpni is set before we get here. */
2002
2003         /* find local NI */
2004         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2005         if (!sd->sd_best_ni) {
2006                 CERROR("Can't send to %s: src %s is not a "
2007                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2008                                 libcfs_nid2str(sd->sd_src_nid));
2009                 return -EINVAL;
2010         }
2011
2012         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2013
2014         return lnet_handle_send(sd);
2015 }
2016
2017 /*
2018  * Source Specified
2019  * Local Destination
2020  * MR Peer
2021  *
2022  * Don't run the selection algorithm on the peer NIs. By specifying the
2023  * local NID, we're also saying that we should always use the destination NID
2024  * provided. This handles the case where we should be using the same
2025  * destination NID for the all the messages which belong to the same RPC
2026  * request.
2027  */
2028 static int
2029 lnet_handle_spec_local_mr_dst(struct lnet_send_data *sd)
2030 {
2031         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2032         if (!sd->sd_best_ni) {
2033                 CERROR("Can't send to %s: src %s is not a "
2034                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2035                                 libcfs_nid2str(sd->sd_src_nid));
2036                 return -EINVAL;
2037         }
2038
2039         if (sd->sd_best_lpni &&
2040             nid_same(&sd->sd_best_lpni->lpni_nid,
2041                       &the_lnet.ln_loni->ni_nid))
2042                 return lnet_handle_lo_send(sd);
2043         else if (sd->sd_best_lpni)
2044                 return lnet_handle_send(sd);
2045
2046         CERROR("can't send to %s. no NI on %s\n",
2047                libcfs_nid2str(sd->sd_dst_nid),
2048                libcfs_net2str(sd->sd_best_ni->ni_net->net_id));
2049
2050         return -EHOSTUNREACH;
2051 }
2052
2053 struct lnet_ni *
2054 lnet_find_best_ni_on_spec_net(struct lnet_ni *cur_best_ni,
2055                               struct lnet_peer *peer,
2056                               struct lnet_peer_net *peer_net,
2057                               struct lnet_msg *msg,
2058                               int cpt)
2059 {
2060         struct lnet_net *local_net;
2061         struct lnet_ni *best_ni;
2062
2063         local_net = lnet_get_net_locked(peer_net->lpn_net_id);
2064         if (!local_net)
2065                 return NULL;
2066
2067         /*
2068          * Iterate through the NIs in this local Net and select
2069          * the NI to send from. The selection is determined by
2070          * these 3 criterion in the following priority:
2071          *      1. NUMA
2072          *      2. NI available credits
2073          *      3. Round Robin
2074          */
2075         best_ni = lnet_get_best_ni(local_net, cur_best_ni,
2076                                    peer, peer_net, msg, cpt);
2077
2078         return best_ni;
2079 }
2080
2081 static int
2082 lnet_initiate_peer_discovery(struct lnet_peer_ni *lpni, struct lnet_msg *msg,
2083                              int cpt)
2084 {
2085         struct lnet_peer *peer;
2086         struct lnet_peer_ni *new_lpni;
2087         int rc;
2088
2089         lnet_peer_ni_addref_locked(lpni);
2090
2091         peer = lpni->lpni_peer_net->lpn_peer;
2092
2093         if (lnet_peer_gw_discovery(peer)) {
2094                 lnet_peer_ni_decref_locked(lpni);
2095                 return 0;
2096         }
2097
2098         if (!lnet_msg_discovery(msg) || lnet_peer_is_uptodate(peer)) {
2099                 lnet_peer_ni_decref_locked(lpni);
2100                 return 0;
2101         }
2102
2103         rc = lnet_discover_peer_locked(lpni, cpt, false);
2104         if (rc) {
2105                 lnet_peer_ni_decref_locked(lpni);
2106                 return rc;
2107         }
2108
2109         new_lpni = lnet_find_peer_ni_locked(lnet_nid_to_nid4(&lpni->lpni_nid));
2110         if (!new_lpni) {
2111                 lnet_peer_ni_decref_locked(lpni);
2112                 return -ENOENT;
2113         }
2114
2115         peer = new_lpni->lpni_peer_net->lpn_peer;
2116         spin_lock(&peer->lp_lock);
2117         if (lpni == new_lpni && lnet_peer_is_uptodate_locked(peer)) {
2118                 /* The peer NI did not change and the peer is up to date.
2119                  * Nothing more to do.
2120                  */
2121                 spin_unlock(&peer->lp_lock);
2122                 lnet_peer_ni_decref_locked(lpni);
2123                 lnet_peer_ni_decref_locked(new_lpni);
2124                 return 0;
2125         }
2126         spin_unlock(&peer->lp_lock);
2127
2128         /* Either the peer NI changed during discovery, or the peer isn't up
2129          * to date. In both cases we want to queue the message on the
2130          * (possibly new) peer's pending queue and queue the peer for discovery
2131          */
2132         msg->msg_sending = 0;
2133         msg->msg_txpeer = NULL;
2134         lnet_net_unlock(cpt);
2135         lnet_peer_queue_message(peer, msg);
2136         lnet_net_lock(cpt);
2137
2138         lnet_peer_ni_decref_locked(lpni);
2139         lnet_peer_ni_decref_locked(new_lpni);
2140
2141         CDEBUG(D_NET, "msg %p delayed. %s pending discovery\n",
2142                msg, libcfs_nidstr(&peer->lp_primary_nid));
2143
2144         return LNET_DC_WAIT;
2145 }
2146
2147 static int
2148 lnet_handle_find_routed_path(struct lnet_send_data *sd,
2149                              lnet_nid_t dst_nid,
2150                              struct lnet_peer_ni **gw_lpni,
2151                              struct lnet_peer **gw_peer)
2152 {
2153         int rc;
2154         struct lnet_peer *gw;
2155         struct lnet_peer *lp;
2156         struct lnet_peer_net *lpn;
2157         struct lnet_peer_net *best_lpn = NULL;
2158         struct lnet_remotenet *rnet, *best_rnet = NULL;
2159         struct lnet_route *best_route = NULL;
2160         struct lnet_route *last_route = NULL;
2161         struct lnet_peer_ni *lpni = NULL;
2162         struct lnet_peer_ni *gwni = NULL;
2163         bool route_found = false;
2164         lnet_nid_t src_nid = (sd->sd_src_nid != LNET_NID_ANY) ? sd->sd_src_nid :
2165                 (sd->sd_best_ni != NULL)
2166                 ? lnet_nid_to_nid4(&sd->sd_best_ni->ni_nid)
2167                 : LNET_NID_ANY;
2168         int best_lpn_healthv = 0;
2169         __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2170
2171         CDEBUG(D_NET, "using src nid %s for route restriction\n",
2172                libcfs_nid2str(src_nid));
2173
2174         /* If a router nid was specified then we are replying to a GET or
2175          * sending an ACK. In this case we use the gateway associated with the
2176          * specified router nid.
2177          */
2178         if (sd->sd_rtr_nid != LNET_NID_ANY) {
2179                 gwni = lnet_find_peer_ni_locked(sd->sd_rtr_nid);
2180                 if (gwni) {
2181                         gw = gwni->lpni_peer_net->lpn_peer;
2182                         lnet_peer_ni_decref_locked(gwni);
2183                         if (gw->lp_rtr_refcount)
2184                                 route_found = true;
2185                 } else {
2186                         CWARN("No peer NI for gateway %s. Attempting to find an alternative route.\n",
2187                                libcfs_nid2str(sd->sd_rtr_nid));
2188                 }
2189         }
2190
2191         if (!route_found) {
2192                 if (sd->sd_msg->msg_routing || src_nid != LNET_NID_ANY) {
2193                         /* If I'm routing this message then I need to find the
2194                          * next hop based on the destination NID
2195                          *
2196                          * We also find next hop based on the destination NID
2197                          * if the source NI was specified
2198                          */
2199                         best_rnet = lnet_find_rnet_locked(LNET_NIDNET(sd->sd_dst_nid));
2200                         if (!best_rnet) {
2201                                 CERROR("Unable to send message from %s to %s - Route table may be misconfigured\n",
2202                                        src_nid != LNET_NID_ANY ?
2203                                                 libcfs_nid2str(src_nid) :
2204                                                 "any local NI",
2205                                        libcfs_nid2str(sd->sd_dst_nid));
2206                                 return -EHOSTUNREACH;
2207                         }
2208                 } else {
2209                         /* we've already looked up the initial lpni using
2210                          * dst_nid
2211                          */
2212                         lpni = sd->sd_best_lpni;
2213                         /* the peer tree must be in existence */
2214                         LASSERT(lpni && lpni->lpni_peer_net &&
2215                                 lpni->lpni_peer_net->lpn_peer);
2216                         lp = lpni->lpni_peer_net->lpn_peer;
2217
2218                         list_for_each_entry(lpn, &lp->lp_peer_nets, lpn_peer_nets) {
2219                                 /* is this remote network reachable?  */
2220                                 rnet = lnet_find_rnet_locked(lpn->lpn_net_id);
2221                                 if (!rnet)
2222                                         continue;
2223
2224                                 if (!best_lpn) {
2225                                         best_lpn = lpn;
2226                                         best_rnet = rnet;
2227                                 }
2228
2229                                 /* select the preferred peer net */
2230                                 if (best_lpn_healthv > lpn->lpn_healthv)
2231                                         continue;
2232                                 else if (best_lpn_healthv < lpn->lpn_healthv)
2233                                         goto use_lpn;
2234
2235                                 if (best_lpn_sel_prio < lpn->lpn_sel_priority)
2236                                         continue;
2237                                 else if (best_lpn_sel_prio > lpn->lpn_sel_priority)
2238                                         goto use_lpn;
2239
2240                                 if (best_lpn->lpn_seq <= lpn->lpn_seq)
2241                                         continue;
2242 use_lpn:
2243                                 best_lpn_healthv = lpn->lpn_healthv;
2244                                 best_lpn_sel_prio = lpn->lpn_sel_priority;
2245                                 best_lpn = lpn;
2246                                 best_rnet = rnet;
2247                         }
2248
2249                         if (!best_lpn) {
2250                                 CERROR("peer %s has no available nets\n",
2251                                        libcfs_nid2str(sd->sd_dst_nid));
2252                                 return -EHOSTUNREACH;
2253                         }
2254
2255                         sd->sd_best_lpni = lnet_find_best_lpni(sd->sd_best_ni,
2256                                                                sd->sd_dst_nid,
2257                                                                lp,
2258                                                                best_lpn->lpn_net_id);
2259                         if (!sd->sd_best_lpni) {
2260                                 CERROR("peer %s is unreachable\n",
2261                                        libcfs_nid2str(sd->sd_dst_nid));
2262                                 return -EHOSTUNREACH;
2263                         }
2264
2265                         /* We're attempting to round robin over the remote peer
2266                          * NI's so update the final destination we selected
2267                          */
2268                         sd->sd_final_dst_lpni = sd->sd_best_lpni;
2269
2270                         /* Increment the sequence number of the remote lpni so
2271                          * we can round robin over the different interfaces of
2272                          * the remote lpni
2273                          */
2274                         sd->sd_best_lpni->lpni_seq++;
2275                 }
2276
2277                 /*
2278                  * find the best route. Restrict the selection on the net of the
2279                  * local NI if we've already picked the local NI to send from.
2280                  * Otherwise, let's pick any route we can find and then find
2281                  * a local NI we can reach the route's gateway on. Any route we
2282                  * select will be reachable by virtue of the restriction we have
2283                  * when adding a route.
2284                  */
2285                 best_route = lnet_find_route_locked(best_rnet,
2286                                                     LNET_NIDNET(src_nid),
2287                                                     sd->sd_best_lpni,
2288                                                     &last_route, &gwni);
2289
2290                 if (!best_route) {
2291                         CERROR("no route to %s from %s\n",
2292                                libcfs_nid2str(dst_nid),
2293                                libcfs_nid2str(src_nid));
2294                         return -EHOSTUNREACH;
2295                 }
2296
2297                 if (!gwni) {
2298                         CERROR("Internal Error. Route expected to %s from %s\n",
2299                                libcfs_nid2str(dst_nid),
2300                                libcfs_nid2str(src_nid));
2301                         return -EFAULT;
2302                 }
2303
2304                 gw = best_route->lr_gateway;
2305                 LASSERT(gw == gwni->lpni_peer_net->lpn_peer);
2306         }
2307
2308         /*
2309          * Discover this gateway if it hasn't already been discovered.
2310          * This means we might delay the message until discovery has
2311          * completed
2312          */
2313         rc = lnet_initiate_peer_discovery(gwni, sd->sd_msg, sd->sd_cpt);
2314         if (rc)
2315                 return rc;
2316
2317         if (!sd->sd_best_ni) {
2318                 lpn = gwni->lpni_peer_net;
2319                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL, gw, lpn,
2320                                                                sd->sd_msg,
2321                                                                sd->sd_md_cpt);
2322                 if (!sd->sd_best_ni) {
2323                         CERROR("Internal Error. Expected local ni on %s but non found: %s\n",
2324                                libcfs_net2str(lpn->lpn_net_id),
2325                                libcfs_nid2str(sd->sd_src_nid));
2326                         return -EFAULT;
2327                 }
2328         }
2329
2330         *gw_lpni = gwni;
2331         *gw_peer = gw;
2332
2333         /*
2334          * increment the sequence numbers since now we're sure we're
2335          * going to use this path
2336          */
2337         if (sd->sd_rtr_nid == LNET_NID_ANY) {
2338                 LASSERT(best_route && last_route);
2339                 best_route->lr_seq = last_route->lr_seq + 1;
2340                 if (best_lpn)
2341                         best_lpn->lpn_seq++;
2342         }
2343
2344         return 0;
2345 }
2346
2347 /*
2348  * Handle two cases:
2349  *
2350  * Case 1:
2351  *  Source specified
2352  *  Remote destination
2353  *  Non-MR destination
2354  *
2355  * Case 2:
2356  *  Source specified
2357  *  Remote destination
2358  *  MR destination
2359  *
2360  * The handling of these two cases is similar. Even though the destination
2361  * can be MR or non-MR, we'll deal directly with the router.
2362  */
2363 static int
2364 lnet_handle_spec_router_dst(struct lnet_send_data *sd)
2365 {
2366         int rc;
2367         struct lnet_peer_ni *gw_lpni = NULL;
2368         struct lnet_peer *gw_peer = NULL;
2369
2370         /* find local NI */
2371         sd->sd_best_ni = lnet_nid2ni_locked(sd->sd_src_nid, sd->sd_cpt);
2372         if (!sd->sd_best_ni) {
2373                 CERROR("Can't send to %s: src %s is not a "
2374                        "local nid\n", libcfs_nid2str(sd->sd_dst_nid),
2375                                 libcfs_nid2str(sd->sd_src_nid));
2376                 return -EINVAL;
2377         }
2378
2379         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2380                                      &gw_peer);
2381         if (rc)
2382                 return rc;
2383
2384         if (sd->sd_send_case & NMR_DST)
2385                 /*
2386                  * since the final destination is non-MR let's set its preferred
2387                  * NID before we send
2388                  */
2389                 lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni,
2390                                          sd->sd_msg);
2391
2392         /*
2393          * We're going to send to the gw found so let's set its
2394          * info
2395          */
2396         sd->sd_peer = gw_peer;
2397         sd->sd_best_lpni = gw_lpni;
2398
2399         return lnet_handle_send(sd);
2400 }
2401
2402 struct lnet_ni *
2403 lnet_find_best_ni_on_local_net(struct lnet_peer *peer, int md_cpt,
2404                                struct lnet_msg *msg, bool discovery)
2405 {
2406         struct lnet_peer_net *lpn = NULL;
2407         struct lnet_peer_net *best_lpn = NULL;
2408         struct lnet_net *net = NULL;
2409         struct lnet_net *best_net = NULL;
2410         struct lnet_ni *best_ni = NULL;
2411         int best_lpn_healthv = 0;
2412         int best_net_healthv = 0;
2413         int net_healthv;
2414         __u32 best_lpn_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2415         __u32 lpn_sel_prio;
2416         __u32 best_net_sel_prio = LNET_MAX_SELECTION_PRIORITY;
2417         __u32 net_sel_prio;
2418         bool exit = false;
2419
2420         /*
2421          * The peer can have multiple interfaces, some of them can be on
2422          * the local network and others on a routed network. We should
2423          * prefer the local network. However if the local network is not
2424          * available then we need to try the routed network
2425          */
2426
2427         /* go through all the peer nets and find the best_ni */
2428         list_for_each_entry(lpn, &peer->lp_peer_nets, lpn_peer_nets) {
2429                 /*
2430                  * The peer's list of nets can contain non-local nets. We
2431                  * want to only examine the local ones.
2432                  */
2433                 net = lnet_get_net_locked(lpn->lpn_net_id);
2434                 if (!net)
2435                         continue;
2436
2437                 lpn_sel_prio = lpn->lpn_sel_priority;
2438                 net_healthv = lnet_get_net_healthv_locked(net);
2439                 net_sel_prio = net->net_sel_priority;
2440
2441                 /*
2442                  * if this is a discovery message and lp_disc_net_id is
2443                  * specified then use that net to send the discovery on.
2444                  */
2445                 if (peer->lp_disc_net_id == lpn->lpn_net_id &&
2446                     discovery) {
2447                         exit = true;
2448                         goto select_lpn;
2449                 }
2450
2451                 if (!best_lpn)
2452                         goto select_lpn;
2453
2454                 /* always select the lpn with the best health */
2455                 if (best_lpn_healthv > lpn->lpn_healthv)
2456                         continue;
2457                 else if (best_lpn_healthv < lpn->lpn_healthv)
2458                         goto select_lpn;
2459
2460                 /* select the preferred peer and local nets */
2461                 if (best_lpn_sel_prio < lpn_sel_prio)
2462                         continue;
2463                 else if (best_lpn_sel_prio > lpn_sel_prio)
2464                         goto select_lpn;
2465
2466                 if (best_net_healthv > net_healthv)
2467                         continue;
2468                 else if (best_net_healthv < net_healthv)
2469                         goto select_lpn;
2470
2471                 if (best_net_sel_prio < net_sel_prio)
2472                         continue;
2473                 else if (best_net_sel_prio > net_sel_prio)
2474                         goto select_lpn;
2475
2476                 if (best_lpn->lpn_seq < lpn->lpn_seq)
2477                         continue;
2478                 else if (best_lpn->lpn_seq > lpn->lpn_seq)
2479                         goto select_lpn;
2480
2481                 /* round robin over the local networks */
2482                 if (best_net->net_seq <= net->net_seq)
2483                         continue;
2484
2485 select_lpn:
2486                 best_net_healthv = net_healthv;
2487                 best_net_sel_prio = net_sel_prio;
2488                 best_lpn_healthv = lpn->lpn_healthv;
2489                 best_lpn_sel_prio = lpn_sel_prio;
2490                 best_lpn = lpn;
2491                 best_net = net;
2492
2493                 if (exit)
2494                         break;
2495         }
2496
2497         if (best_lpn) {
2498                 /* Select the best NI on the same net as best_lpn chosen
2499                  * above
2500                  */
2501                 best_ni = lnet_find_best_ni_on_spec_net(NULL, peer, best_lpn,
2502                                                         msg, md_cpt);
2503         }
2504
2505         return best_ni;
2506 }
2507
2508 static struct lnet_ni *
2509 lnet_find_existing_preferred_best_ni(struct lnet_peer_ni *lpni, int cpt)
2510 {
2511         struct lnet_ni *best_ni = NULL;
2512         struct lnet_peer_net *peer_net = lpni->lpni_peer_net;
2513         struct lnet_peer_ni *lpni_entry;
2514
2515         /*
2516          * We must use a consistent source address when sending to a
2517          * non-MR peer. However, a non-MR peer can have multiple NIDs
2518          * on multiple networks, and we may even need to talk to this
2519          * peer on multiple networks -- certain types of
2520          * load-balancing configuration do this.
2521          *
2522          * So we need to pick the NI the peer prefers for this
2523          * particular network.
2524          */
2525         LASSERT(peer_net);
2526         list_for_each_entry(lpni_entry, &peer_net->lpn_peer_nis,
2527                             lpni_peer_nis) {
2528                 if (lpni_entry->lpni_pref_nnids == 0)
2529                         continue;
2530                 LASSERT(lpni_entry->lpni_pref_nnids == 1);
2531                 best_ni = lnet_nid_to_ni_locked(&lpni_entry->lpni_pref.nid,
2532                                                 cpt);
2533                 break;
2534         }
2535
2536         return best_ni;
2537 }
2538
2539 /* Prerequisite: sd->sd_peer and sd->sd_best_lpni should be set */
2540 static int
2541 lnet_select_preferred_best_ni(struct lnet_send_data *sd)
2542 {
2543         struct lnet_ni *best_ni = NULL;
2544         struct lnet_peer_ni *best_lpni = sd->sd_best_lpni;
2545
2546         /*
2547          * We must use a consistent source address when sending to a
2548          * non-MR peer. However, a non-MR peer can have multiple NIDs
2549          * on multiple networks, and we may even need to talk to this
2550          * peer on multiple networks -- certain types of
2551          * load-balancing configuration do this.
2552          *
2553          * So we need to pick the NI the peer prefers for this
2554          * particular network.
2555          */
2556
2557         best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2558                                                        sd->sd_cpt);
2559
2560         /* if best_ni is still not set just pick one */
2561         if (!best_ni) {
2562                 best_ni =
2563                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2564                                                 sd->sd_best_lpni->lpni_peer_net,
2565                                                 sd->sd_msg,
2566                                                 sd->sd_md_cpt);
2567                 /* If there is no best_ni we don't have a route */
2568                 if (!best_ni) {
2569                         CERROR("no path to %s from net %s\n",
2570                                 libcfs_nidstr(&best_lpni->lpni_nid),
2571                                 libcfs_net2str(best_lpni->lpni_net->net_id));
2572                         return -EHOSTUNREACH;
2573                 }
2574         }
2575
2576         sd->sd_best_ni = best_ni;
2577
2578         /* Set preferred NI if necessary. */
2579         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2580
2581         return 0;
2582 }
2583
2584
2585 /*
2586  * Source not specified
2587  * Local destination
2588  * Non-MR Peer
2589  *
2590  * always use the same source NID for NMR peers
2591  * If we've talked to that peer before then we already have a preferred
2592  * source NI associated with it. Otherwise, we select a preferred local NI
2593  * and store it in the peer
2594  */
2595 static int
2596 lnet_handle_any_local_nmr_dst(struct lnet_send_data *sd)
2597 {
2598         int rc = 0;
2599
2600         /* sd->sd_best_lpni is already set to the final destination */
2601
2602         /*
2603          * At this point we should've created the peer ni and peer. If we
2604          * can't find it, then something went wrong. Instead of assert
2605          * output a relevant message and fail the send
2606          */
2607         if (!sd->sd_best_lpni) {
2608                 CERROR("Internal fault. Unable to send msg %s to %s. "
2609                        "NID not known\n",
2610                        lnet_msgtyp2str(sd->sd_msg->msg_type),
2611                        libcfs_nid2str(sd->sd_dst_nid));
2612                 return -EFAULT;
2613         }
2614
2615         if (sd->sd_msg->msg_routing) {
2616                 /* If I'm forwarding this message then I can choose any NI
2617                  * on the destination peer net
2618                  */
2619                 sd->sd_best_ni = lnet_find_best_ni_on_spec_net(NULL,
2620                                                                sd->sd_peer,
2621                                                                sd->sd_best_lpni->lpni_peer_net,
2622                                                                sd->sd_msg,
2623                                                                sd->sd_md_cpt);
2624                 if (!sd->sd_best_ni) {
2625                         CERROR("Unable to forward message to %s. No local NI available\n",
2626                                libcfs_nid2str(sd->sd_dst_nid));
2627                         rc = -EHOSTUNREACH;
2628                 }
2629         } else
2630                 rc = lnet_select_preferred_best_ni(sd);
2631
2632         if (!rc)
2633                 rc = lnet_handle_send(sd);
2634
2635         return rc;
2636 }
2637
2638 static int
2639 lnet_handle_any_mr_dsta(struct lnet_send_data *sd)
2640 {
2641         /*
2642          * NOTE we've already handled the remote peer case. So we only
2643          * need to worry about the local case here.
2644          *
2645          * if we're sending a response, ACK or reply, we need to send it
2646          * to the destination NID given to us. At this point we already
2647          * have the peer_ni we're suppose to send to, so just find the
2648          * best_ni on the peer net and use that. Since we're sending to an
2649          * MR peer then we can just run the selection algorithm on our
2650          * local NIs and pick the best one.
2651          */
2652         if (sd->sd_send_case & SND_RESP) {
2653                 sd->sd_best_ni =
2654                   lnet_find_best_ni_on_spec_net(NULL, sd->sd_peer,
2655                                                 sd->sd_best_lpni->lpni_peer_net,
2656                                                 sd->sd_msg,
2657                                                 sd->sd_md_cpt);
2658
2659                 if (!sd->sd_best_ni) {
2660                         /*
2661                          * We're not going to deal with not able to send
2662                          * a response to the provided final destination
2663                          */
2664                         CERROR("Can't send response to %s. "
2665                                "No local NI available\n",
2666                                 libcfs_nid2str(sd->sd_dst_nid));
2667                         return -EHOSTUNREACH;
2668                 }
2669
2670                 return lnet_handle_send(sd);
2671         }
2672
2673         /*
2674          * If we get here that means we're sending a fresh request, PUT or
2675          * GET, so we need to run our standard selection algorithm.
2676          * First find the best local interface that's on any of the peer's
2677          * networks.
2678          */
2679         sd->sd_best_ni = lnet_find_best_ni_on_local_net(sd->sd_peer,
2680                                         sd->sd_md_cpt,
2681                                         sd->sd_msg,
2682                                         lnet_msg_discovery(sd->sd_msg));
2683         if (sd->sd_best_ni) {
2684                 sd->sd_best_lpni =
2685                   lnet_find_best_lpni(sd->sd_best_ni, sd->sd_dst_nid,
2686                                       sd->sd_peer,
2687                                       sd->sd_best_ni->ni_net->net_id);
2688
2689                 /*
2690                  * if we're successful in selecting a peer_ni on the local
2691                  * network, then send to it. Otherwise fall through and
2692                  * try and see if we can reach it over another routed
2693                  * network
2694                  */
2695                 if (sd->sd_best_lpni &&
2696                     nid_same(&sd->sd_best_lpni->lpni_nid,
2697                              &the_lnet.ln_loni->ni_nid)) {
2698                         /*
2699                          * in case we initially started with a routed
2700                          * destination, let's reset to local
2701                          */
2702                         sd->sd_send_case &= ~REMOTE_DST;
2703                         sd->sd_send_case |= LOCAL_DST;
2704                         return lnet_handle_lo_send(sd);
2705                 } else if (sd->sd_best_lpni) {
2706                         /*
2707                          * in case we initially started with a routed
2708                          * destination, let's reset to local
2709                          */
2710                         sd->sd_send_case &= ~REMOTE_DST;
2711                         sd->sd_send_case |= LOCAL_DST;
2712                         return lnet_handle_send(sd);
2713                 }
2714
2715                 CERROR("Internal Error. Expected to have a best_lpni: "
2716                        "%s -> %s\n",
2717                        libcfs_nid2str(sd->sd_src_nid),
2718                        libcfs_nid2str(sd->sd_dst_nid));
2719
2720                 return -EFAULT;
2721         }
2722
2723         /*
2724          * Peer doesn't have a local network. Let's see if there is
2725          * a remote network we can reach it on.
2726          */
2727         return PASS_THROUGH;
2728 }
2729
2730 /*
2731  * Case 1:
2732  *      Source NID not specified
2733  *      Local destination
2734  *      MR peer
2735  *
2736  * Case 2:
2737  *      Source NID not speified
2738  *      Remote destination
2739  *      MR peer
2740  *
2741  * In both of these cases if we're sending a response, ACK or REPLY, then
2742  * we need to send to the destination NID provided.
2743  *
2744  * In the remote case let's deal with MR routers.
2745  *
2746  */
2747
2748 static int
2749 lnet_handle_any_mr_dst(struct lnet_send_data *sd)
2750 {
2751         int rc = 0;
2752         struct lnet_peer *gw_peer = NULL;
2753         struct lnet_peer_ni *gw_lpni = NULL;
2754
2755         /*
2756          * handle sending a response to a remote peer here so we don't
2757          * have to worry about it if we hit lnet_handle_any_mr_dsta()
2758          */
2759         if (sd->sd_send_case & REMOTE_DST &&
2760             sd->sd_send_case & SND_RESP) {
2761                 struct lnet_peer_ni *gw;
2762                 struct lnet_peer *gw_peer;
2763
2764                 rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw,
2765                                                   &gw_peer);
2766                 if (rc < 0) {
2767                         CERROR("Can't send response to %s. "
2768                                "No route available\n",
2769                                 libcfs_nid2str(sd->sd_dst_nid));
2770                         return -EHOSTUNREACH;
2771                 } else if (rc > 0) {
2772                         return rc;
2773                 }
2774
2775                 sd->sd_best_lpni = gw;
2776                 sd->sd_peer = gw_peer;
2777
2778                 return lnet_handle_send(sd);
2779         }
2780
2781         /*
2782          * Even though the NID for the peer might not be on a local network,
2783          * since the peer is MR there could be other interfaces on the
2784          * local network. In that case we'd still like to prefer the local
2785          * network over the routed network. If we're unable to do that
2786          * then we select the best router among the different routed networks,
2787          * and if the router is MR then we can deal with it as such.
2788          */
2789         rc = lnet_handle_any_mr_dsta(sd);
2790         if (rc != PASS_THROUGH)
2791                 return rc;
2792
2793         /*
2794          * Now that we must route to the destination, we must consider the
2795          * MR case, where the destination has multiple interfaces, some of
2796          * which we can route to and others we do not. For this reason we
2797          * need to select the destination which we can route to and if
2798          * there are multiple, we need to round robin.
2799          */
2800         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2801                                           &gw_peer);
2802         if (rc)
2803                 return rc;
2804
2805         sd->sd_send_case &= ~LOCAL_DST;
2806         sd->sd_send_case |= REMOTE_DST;
2807
2808         sd->sd_peer = gw_peer;
2809         sd->sd_best_lpni = gw_lpni;
2810
2811         return lnet_handle_send(sd);
2812 }
2813
2814 /*
2815  * Source not specified
2816  * Remote destination
2817  * Non-MR peer
2818  *
2819  * Must send to the specified peer NID using the same source NID that
2820  * we've used before. If it's the first time to talk to that peer then
2821  * find the source NI and assign it as preferred to that peer
2822  */
2823 static int
2824 lnet_handle_any_router_nmr_dst(struct lnet_send_data *sd)
2825 {
2826         int rc;
2827         struct lnet_peer_ni *gw_lpni = NULL;
2828         struct lnet_peer *gw_peer = NULL;
2829
2830         /*
2831          * Let's see if we have a preferred NI to talk to this NMR peer
2832          */
2833         sd->sd_best_ni = lnet_find_existing_preferred_best_ni(sd->sd_best_lpni,
2834                                                               sd->sd_cpt);
2835
2836         /*
2837          * find the router and that'll find the best NI if we didn't find
2838          * it already.
2839          */
2840         rc = lnet_handle_find_routed_path(sd, sd->sd_dst_nid, &gw_lpni,
2841                                           &gw_peer);
2842         if (rc)
2843                 return rc;
2844
2845         /*
2846          * set the best_ni we've chosen as the preferred one for
2847          * this peer
2848          */
2849         lnet_set_non_mr_pref_nid(sd->sd_best_lpni, sd->sd_best_ni, sd->sd_msg);
2850
2851         /* we'll be sending to the gw */
2852         sd->sd_best_lpni = gw_lpni;
2853         sd->sd_peer = gw_peer;
2854
2855         return lnet_handle_send(sd);
2856 }
2857
2858 static int
2859 lnet_handle_send_case_locked(struct lnet_send_data *sd)
2860 {
2861         /*
2862          * turn off the SND_RESP bit.
2863          * It will be checked in the case handling
2864          */
2865         __u32 send_case = sd->sd_send_case &= ~SND_RESP ;
2866
2867         CDEBUG(D_NET, "Source %s%s to %s %s %s destination\n",
2868                 (send_case & SRC_SPEC) ? "Specified: " : "ANY",
2869                 (send_case & SRC_SPEC) ? libcfs_nid2str(sd->sd_src_nid) : "",
2870                 (send_case & MR_DST) ? "MR: " : "NMR: ",
2871                 libcfs_nid2str(sd->sd_dst_nid),
2872                 (send_case & LOCAL_DST) ? "local" : "routed");
2873
2874         switch (send_case) {
2875         /*
2876          * For all cases where the source is specified, we should always
2877          * use the destination NID, whether it's an MR destination or not,
2878          * since we're continuing a series of related messages for the
2879          * same RPC
2880          */
2881         case SRC_SPEC_LOCAL_NMR_DST:
2882                 return lnet_handle_spec_local_nmr_dst(sd);
2883         case SRC_SPEC_LOCAL_MR_DST:
2884                 return lnet_handle_spec_local_mr_dst(sd);
2885         case SRC_SPEC_ROUTER_NMR_DST:
2886         case SRC_SPEC_ROUTER_MR_DST:
2887                 return lnet_handle_spec_router_dst(sd);
2888         case SRC_ANY_LOCAL_NMR_DST:
2889                 return lnet_handle_any_local_nmr_dst(sd);
2890         case SRC_ANY_LOCAL_MR_DST:
2891         case SRC_ANY_ROUTER_MR_DST:
2892                 return lnet_handle_any_mr_dst(sd);
2893         case SRC_ANY_ROUTER_NMR_DST:
2894                 return lnet_handle_any_router_nmr_dst(sd);
2895         default:
2896                 CERROR("Unknown send case\n");
2897                 return -1;
2898         }
2899 }
2900
2901 static int
2902 lnet_select_pathway(lnet_nid_t src_nid, lnet_nid_t dst_nid,
2903                     struct lnet_msg *msg, lnet_nid_t rtr_nid)
2904 {
2905         struct lnet_peer_ni *lpni;
2906         struct lnet_peer *peer;
2907         struct lnet_send_data send_data;
2908         int cpt, rc;
2909         int md_cpt;
2910         __u32 send_case = 0;
2911         bool final_hop;
2912         bool mr_forwarding_allowed;
2913
2914         memset(&send_data, 0, sizeof(send_data));
2915
2916         /*
2917          * get an initial CPT to use for locking. The idea here is not to
2918          * serialize the calls to select_pathway, so that as many
2919          * operations can run concurrently as possible. To do that we use
2920          * the CPT where this call is being executed. Later on when we
2921          * determine the CPT to use in lnet_message_commit, we switch the
2922          * lock and check if there was any configuration change.  If none,
2923          * then we proceed, if there is, then we restart the operation.
2924          */
2925         cpt = lnet_net_lock_current();
2926
2927         md_cpt = lnet_cpt_of_md(msg->msg_md, msg->msg_offset);
2928         if (md_cpt == CFS_CPT_ANY)
2929                 md_cpt = cpt;
2930
2931 again:
2932
2933         /*
2934          * If we're being asked to send to the loopback interface, there
2935          * is no need to go through any selection. We can just shortcut
2936          * the entire process and send over lolnd
2937          */
2938         send_data.sd_msg = msg;
2939         send_data.sd_cpt = cpt;
2940         if (dst_nid == LNET_NID_LO_0) {
2941                 rc = lnet_handle_lo_send(&send_data);
2942                 lnet_net_unlock(cpt);
2943                 return rc;
2944         }
2945
2946         /*
2947          * find an existing peer_ni, or create one and mark it as having been
2948          * created due to network traffic. This call will create the
2949          * peer->peer_net->peer_ni tree.
2950          */
2951         lpni = lnet_nid2peerni_locked(dst_nid, LNET_NID_ANY, cpt);
2952         if (IS_ERR(lpni)) {
2953                 lnet_net_unlock(cpt);
2954                 return PTR_ERR(lpni);
2955         }
2956
2957         /*
2958          * Cache the original src_nid and rtr_nid. If we need to resend the
2959          * message then we'll need to know whether the src_nid was originally
2960          * specified for this message. If it was originally specified,
2961          * then we need to keep using the same src_nid since it's
2962          * continuing the same sequence of messages. Similarly, rtr_nid will
2963          * affect our choice of next hop.
2964          */
2965         msg->msg_src_nid_param = src_nid;
2966         msg->msg_rtr_nid_param = rtr_nid;
2967
2968         /*
2969          * If necessary, perform discovery on the peer that owns this peer_ni.
2970          * Note, this can result in the ownership of this peer_ni changing
2971          * to another peer object.
2972          */
2973         rc = lnet_initiate_peer_discovery(lpni, msg, cpt);
2974         if (rc) {
2975                 lnet_peer_ni_decref_locked(lpni);
2976                 lnet_net_unlock(cpt);
2977                 return rc;
2978         }
2979         lnet_peer_ni_decref_locked(lpni);
2980
2981         peer = lpni->lpni_peer_net->lpn_peer;
2982
2983         /*
2984          * Identify the different send cases
2985          */
2986         if (src_nid == LNET_NID_ANY) {
2987                 send_case |= SRC_ANY;
2988                 if (lnet_get_net_locked(LNET_NIDNET(dst_nid)))
2989                         send_case |= LOCAL_DST;
2990                 else
2991                         send_case |= REMOTE_DST;
2992         } else {
2993                 send_case |= SRC_SPEC;
2994                 if (LNET_NIDNET(src_nid) == LNET_NIDNET(dst_nid))
2995                         send_case |= LOCAL_DST;
2996                 else
2997                         send_case |= REMOTE_DST;
2998         }
2999
3000         final_hop = false;
3001         if (msg->msg_routing && (send_case & LOCAL_DST))
3002                 final_hop = true;
3003
3004         /* Determine whether to allow MR forwarding for this message.
3005          * NB: MR forwarding is allowed if the message originator and the
3006          * destination are both MR capable, and the destination lpni that was
3007          * originally chosen by the originator is unhealthy or down.
3008          * We check the MR capability of the destination further below
3009          */
3010         mr_forwarding_allowed = false;
3011         if (final_hop) {
3012                 struct lnet_peer *src_lp;
3013                 struct lnet_peer_ni *src_lpni;
3014
3015                 src_lpni = lnet_nid2peerni_locked(msg->msg_hdr.src_nid,
3016                                                   LNET_NID_ANY, cpt);
3017                 /* We don't fail the send if we hit any errors here. We'll just
3018                  * try to send it via non-multi-rail criteria
3019                  */
3020                 if (!IS_ERR(src_lpni)) {
3021                         /* Drop ref taken by lnet_nid2peerni_locked() */
3022                         lnet_peer_ni_decref_locked(src_lpni);
3023                         src_lp = lpni->lpni_peer_net->lpn_peer;
3024                         if (lnet_peer_is_multi_rail(src_lp) &&
3025                             !lnet_is_peer_ni_alive(lpni))
3026                                 mr_forwarding_allowed = true;
3027
3028                 }
3029                 CDEBUG(D_NET, "msg %p MR forwarding %s\n", msg,
3030                        mr_forwarding_allowed ? "allowed" : "not allowed");
3031         }
3032
3033         /*
3034          * Deal with the peer as NMR in the following cases:
3035          * 1. the peer is NMR
3036          * 2. We're trying to recover a specific peer NI
3037          * 3. I'm a router sending to the final destination and MR forwarding is
3038          *    not allowed for this message (as determined above).
3039          *    In this case the source of the message would've
3040          *    already selected the final destination so my job
3041          *    is to honor the selection.
3042          */
3043         if (!lnet_peer_is_multi_rail(peer) || msg->msg_recovery ||
3044             (final_hop && !mr_forwarding_allowed))
3045                 send_case |= NMR_DST;
3046         else
3047                 send_case |= MR_DST;
3048
3049         if (lnet_msg_is_response(msg))
3050                 send_case |= SND_RESP;
3051
3052         /* assign parameters to the send_data */
3053         send_data.sd_rtr_nid = rtr_nid;
3054         send_data.sd_src_nid = src_nid;
3055         send_data.sd_dst_nid = dst_nid;
3056         send_data.sd_best_lpni = lpni;
3057         /*
3058          * keep a pointer to the final destination in case we're going to
3059          * route, so we'll need to access it later
3060          */
3061         send_data.sd_final_dst_lpni = lpni;
3062         send_data.sd_peer = peer;
3063         send_data.sd_md_cpt = md_cpt;
3064         send_data.sd_send_case = send_case;
3065
3066         rc = lnet_handle_send_case_locked(&send_data);
3067
3068         /*
3069          * Update the local cpt since send_data.sd_cpt might've been
3070          * updated as a result of calling lnet_handle_send_case_locked().
3071          */
3072         cpt = send_data.sd_cpt;
3073
3074         if (rc == REPEAT_SEND)
3075                 goto again;
3076
3077         lnet_net_unlock(cpt);
3078
3079         return rc;
3080 }
3081
3082 int
3083 lnet_send(lnet_nid_t src_nid, struct lnet_msg *msg, lnet_nid_t rtr_nid)
3084 {
3085         lnet_nid_t              dst_nid = msg->msg_target.nid;
3086         int                     rc;
3087
3088         /*
3089          * NB: rtr_nid is set to LNET_NID_ANY for all current use-cases,
3090          * but we might want to use pre-determined router for ACK/REPLY
3091          * in the future
3092          */
3093         /* NB: ni != NULL == interface pre-determined (ACK/REPLY) */
3094         LASSERT(msg->msg_txpeer == NULL);
3095         LASSERT(msg->msg_txni == NULL);
3096         LASSERT(!msg->msg_sending);
3097         LASSERT(!msg->msg_target_is_router);
3098         LASSERT(!msg->msg_receiving);
3099
3100         msg->msg_sending = 1;
3101
3102         LASSERT(!msg->msg_tx_committed);
3103
3104         rc = lnet_select_pathway(src_nid, dst_nid, msg, rtr_nid);
3105         if (rc < 0) {
3106                 if (rc == -EHOSTUNREACH)
3107                         msg->msg_health_status = LNET_MSG_STATUS_REMOTE_ERROR;
3108                 else
3109                         msg->msg_health_status = LNET_MSG_STATUS_LOCAL_ERROR;
3110                 return rc;
3111         }
3112
3113         if (rc == LNET_CREDIT_OK)
3114                 lnet_ni_send(msg->msg_txni, msg);
3115
3116         /* rc == LNET_CREDIT_OK or LNET_CREDIT_WAIT or LNET_DC_WAIT */
3117         return 0;
3118 }
3119
3120 enum lnet_mt_event_type {
3121         MT_TYPE_LOCAL_NI = 0,
3122         MT_TYPE_PEER_NI
3123 };
3124
3125 struct lnet_mt_event_info {
3126         enum lnet_mt_event_type mt_type;
3127         lnet_nid_t mt_nid;
3128 };
3129
3130 /* called with res_lock held */
3131 void
3132 lnet_detach_rsp_tracker(struct lnet_libmd *md, int cpt)
3133 {
3134         struct lnet_rsp_tracker *rspt;
3135
3136         /*
3137          * msg has a refcount on the MD so the MD is not going away.
3138          * The rspt queue for the cpt is protected by
3139          * the lnet_net_lock(cpt). cpt is the cpt of the MD cookie.
3140          */
3141         if (!md->md_rspt_ptr)
3142                 return;
3143
3144         rspt = md->md_rspt_ptr;
3145
3146         /* debug code */
3147         LASSERT(rspt->rspt_cpt == cpt);
3148
3149         md->md_rspt_ptr = NULL;
3150
3151         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3152                 /*
3153                  * The monitor thread has invalidated this handle because the
3154                  * response timed out, but it failed to lookup the MD. That
3155                  * means this response tracker is on the zombie list. We can
3156                  * safely remove it under the resource lock (held by caller) and
3157                  * free the response tracker block.
3158                  */
3159                 list_del(&rspt->rspt_on_list);
3160                 lnet_rspt_free(rspt, cpt);
3161         } else {
3162                 /*
3163                  * invalidate the handle to indicate that a response has been
3164                  * received, which will then lead the monitor thread to clean up
3165                  * the rspt block.
3166                  */
3167                 LNetInvalidateMDHandle(&rspt->rspt_mdh);
3168         }
3169 }
3170
3171 void
3172 lnet_clean_zombie_rstqs(void)
3173 {
3174         struct lnet_rsp_tracker *rspt, *tmp;
3175         int i;
3176
3177         cfs_cpt_for_each(i, lnet_cpt_table()) {
3178                 list_for_each_entry_safe(rspt, tmp,
3179                                          the_lnet.ln_mt_zombie_rstqs[i],
3180                                          rspt_on_list) {
3181                         list_del(&rspt->rspt_on_list);
3182                         lnet_rspt_free(rspt, i);
3183                 }
3184         }
3185
3186         cfs_percpt_free(the_lnet.ln_mt_zombie_rstqs);
3187 }
3188
3189 static void
3190 lnet_finalize_expired_responses(void)
3191 {
3192         struct lnet_libmd *md;
3193         struct lnet_rsp_tracker *rspt, *tmp;
3194         ktime_t now;
3195         int i;
3196
3197         if (the_lnet.ln_mt_rstq == NULL)
3198                 return;
3199
3200         cfs_cpt_for_each(i, lnet_cpt_table()) {
3201                 LIST_HEAD(local_queue);
3202
3203                 lnet_net_lock(i);
3204                 if (!the_lnet.ln_mt_rstq[i]) {
3205                         lnet_net_unlock(i);
3206                         continue;
3207                 }
3208                 list_splice_init(the_lnet.ln_mt_rstq[i], &local_queue);
3209                 lnet_net_unlock(i);
3210
3211                 now = ktime_get();
3212
3213                 list_for_each_entry_safe(rspt, tmp, &local_queue, rspt_on_list) {
3214                         /*
3215                          * The rspt mdh will be invalidated when a response
3216                          * is received or whenever we want to discard the
3217                          * block the monitor thread will walk the queue
3218                          * and clean up any rsts with an invalid mdh.
3219                          * The monitor thread will walk the queue until
3220                          * the first unexpired rspt block. This means that
3221                          * some rspt blocks which received their
3222                          * corresponding responses will linger in the
3223                          * queue until they are cleaned up eventually.
3224                          */
3225                         lnet_res_lock(i);
3226                         if (LNetMDHandleIsInvalid(rspt->rspt_mdh)) {
3227                                 lnet_res_unlock(i);
3228                                 list_del(&rspt->rspt_on_list);
3229                                 lnet_rspt_free(rspt, i);
3230                                 continue;
3231                         }
3232
3233                         if (ktime_compare(now, rspt->rspt_deadline) >= 0 ||
3234                             the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN) {
3235                                 struct lnet_peer_ni *lpni;
3236                                 struct lnet_nid nid;
3237
3238                                 md = lnet_handle2md(&rspt->rspt_mdh);
3239                                 if (!md) {
3240                                         /* MD has been queued for unlink, but
3241                                          * rspt hasn't been detached (Note we've
3242                                          * checked above that the rspt_mdh is
3243                                          * valid). Since we cannot lookup the MD
3244                                          * we're unable to detach the rspt
3245                                          * ourselves. Thus, move the rspt to the
3246                                          * zombie list where we'll wait for
3247                                          * either:
3248                                          *   1. The remaining operations on the
3249                                          *   MD to complete. In this case the
3250                                          *   final operation will result in
3251                                          *   lnet_msg_detach_md()->
3252                                          *   lnet_detach_rsp_tracker() where
3253                                          *   we will clean up this response
3254                                          *   tracker.
3255                                          *   2. LNet to shutdown. In this case
3256                                          *   we'll wait until after all LND Nets
3257                                          *   have shutdown and then we can
3258                                          *   safely free any remaining response
3259                                          *   tracker blocks on the zombie list.
3260                                          * Note: We need to hold the resource
3261                                          * lock when adding to the zombie list
3262                                          * because we may have concurrent access
3263                                          * with lnet_detach_rsp_tracker().
3264                                          */
3265                                         LNetInvalidateMDHandle(&rspt->rspt_mdh);
3266                                         list_move(&rspt->rspt_on_list,
3267                                                   the_lnet.ln_mt_zombie_rstqs[i]);
3268                                         lnet_res_unlock(i);
3269                                         continue;
3270                                 }
3271                                 LASSERT(md->md_rspt_ptr == rspt);
3272                                 md->md_rspt_ptr = NULL;
3273                                 lnet_res_unlock(i);
3274
3275                                 LNetMDUnlink(rspt->rspt_mdh);
3276
3277                                 nid = rspt->rspt_next_hop_nid;
3278
3279                                 list_del(&rspt->rspt_on_list);
3280                                 lnet_rspt_free(rspt, i);
3281
3282                                 /* If we're shutting down we just want to clean
3283                                  * up the rspt blocks
3284                                  */
3285                                 if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
3286                                         continue;
3287
3288                                 lnet_net_lock(i);
3289                                 the_lnet.ln_counters[i]->lct_health.lch_response_timeout_count++;
3290                                 lnet_net_unlock(i);
3291
3292                                 CDEBUG(D_NET,
3293                                        "Response timeout: md = %p: nid = %s\n",
3294                                        md, libcfs_nidstr(&nid));
3295
3296                                 /*
3297                                  * If there is a timeout on the response
3298                                  * from the next hop decrement its health
3299                                  * value so that we don't use it
3300                                  */
3301                                 lnet_net_lock(0);
3302                                 lpni = lnet_peer_ni_find_locked(&nid);
3303                                 if (lpni) {
3304                                         lnet_handle_remote_failure_locked(lpni);
3305                                         lnet_peer_ni_decref_locked(lpni);
3306                                 }
3307                                 lnet_net_unlock(0);
3308                         } else {
3309                                 lnet_res_unlock(i);
3310                                 break;
3311                         }
3312                 }
3313
3314                 if (!list_empty(&local_queue)) {
3315                         lnet_net_lock(i);
3316                         list_splice(&local_queue, the_lnet.ln_mt_rstq[i]);
3317                         lnet_net_unlock(i);
3318                 }
3319         }
3320 }
3321
3322 static void
3323 lnet_resend_pending_msgs_locked(struct list_head *resendq, int cpt)
3324 {
3325         struct lnet_msg *msg;
3326
3327         while (!list_empty(resendq)) {
3328                 struct lnet_peer_ni *lpni;
3329
3330                 msg = list_entry(resendq->next, struct lnet_msg,
3331                                  msg_list);
3332
3333                 list_del_init(&msg->msg_list);
3334
3335                 lpni = lnet_find_peer_ni_locked(msg->msg_hdr.dest_nid);
3336                 if (!lpni) {
3337                         lnet_net_unlock(cpt);
3338                         CERROR("Expected that a peer is already created for %s\n",
3339                                libcfs_nid2str(msg->msg_hdr.dest_nid));
3340                         msg->msg_no_resend = true;
3341                         lnet_finalize(msg, -EFAULT);
3342                         lnet_net_lock(cpt);
3343                 } else {
3344                         int rc;
3345
3346                         lnet_peer_ni_decref_locked(lpni);
3347
3348                         lnet_net_unlock(cpt);
3349                         CDEBUG(D_NET, "resending %s->%s: %s recovery %d try# %d\n",
3350                                libcfs_nid2str(msg->msg_src_nid_param),
3351                                libcfs_id2str(msg->msg_target),
3352                                lnet_msgtyp2str(msg->msg_type),
3353                                msg->msg_recovery,
3354                                msg->msg_retry_count);
3355                         rc = lnet_send(msg->msg_src_nid_param, msg,
3356                                        msg->msg_rtr_nid_param);
3357                         if (rc) {
3358                                 CERROR("Error sending %s to %s: %d\n",
3359                                        lnet_msgtyp2str(msg->msg_type),
3360                                        libcfs_id2str(msg->msg_target), rc);
3361                                 msg->msg_no_resend = true;
3362                                 lnet_finalize(msg, rc);
3363                         }
3364                         lnet_net_lock(cpt);
3365                         if (!rc)
3366                                 the_lnet.ln_counters[cpt]->lct_health.lch_resend_count++;
3367                 }
3368         }
3369 }
3370
3371 static void
3372 lnet_resend_pending_msgs(void)
3373 {
3374         int i;
3375
3376         cfs_cpt_for_each(i, lnet_cpt_table()) {
3377                 lnet_net_lock(i);
3378                 lnet_resend_pending_msgs_locked(the_lnet.ln_mt_resendqs[i], i);
3379                 lnet_net_unlock(i);
3380         }
3381 }
3382
3383 /* called with cpt and ni_lock held */
3384 static void
3385 lnet_unlink_ni_recovery_mdh_locked(struct lnet_ni *ni, int cpt, bool force)
3386 {
3387         struct lnet_handle_md recovery_mdh;
3388
3389         LNetInvalidateMDHandle(&recovery_mdh);
3390
3391         if (ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING ||
3392             force) {
3393                 recovery_mdh = ni->ni_ping_mdh;
3394                 LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3395         }
3396         lnet_ni_unlock(ni);
3397         lnet_net_unlock(cpt);
3398         if (!LNetMDHandleIsInvalid(recovery_mdh))
3399                 LNetMDUnlink(recovery_mdh);
3400         lnet_net_lock(cpt);
3401         lnet_ni_lock(ni);
3402 }
3403
3404 static void
3405 lnet_recover_local_nis(void)
3406 {
3407         struct lnet_mt_event_info *ev_info;
3408         LIST_HEAD(processed_list);
3409         LIST_HEAD(local_queue);
3410         struct lnet_handle_md mdh;
3411         struct lnet_ni *tmp;
3412         struct lnet_ni *ni;
3413         lnet_nid_t nid;
3414         int healthv;
3415         int rc;
3416         time64_t now;
3417
3418         /*
3419          * splice the recovery queue on a local queue. We will iterate
3420          * through the local queue and update it as needed. Once we're
3421          * done with the traversal, we'll splice the local queue back on
3422          * the head of the ln_mt_localNIRecovq. Any newly added local NIs
3423          * will be traversed in the next iteration.
3424          */
3425         lnet_net_lock(0);
3426         list_splice_init(&the_lnet.ln_mt_localNIRecovq,
3427                          &local_queue);
3428         lnet_net_unlock(0);
3429
3430         now = ktime_get_seconds();
3431
3432         list_for_each_entry_safe(ni, tmp, &local_queue, ni_recovery) {
3433                 /*
3434                  * if an NI is being deleted or it is now healthy, there
3435                  * is no need to keep it around in the recovery queue.
3436                  * The monitor thread is the only thread responsible for
3437                  * removing the NI from the recovery queue.
3438                  * Multiple threads can be adding NIs to the recovery
3439                  * queue.
3440                  */
3441                 healthv = atomic_read(&ni->ni_healthv);
3442
3443                 lnet_net_lock(0);
3444                 lnet_ni_lock(ni);
3445                 if (ni->ni_state != LNET_NI_STATE_ACTIVE ||
3446                     healthv == LNET_MAX_HEALTH_VALUE) {
3447                         list_del_init(&ni->ni_recovery);
3448                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, false);
3449                         lnet_ni_unlock(ni);
3450                         lnet_ni_decref_locked(ni, 0);
3451                         lnet_net_unlock(0);
3452                         continue;
3453                 }
3454
3455                 /*
3456                  * if the local NI failed recovery we must unlink the md.
3457                  * But we want to keep the local_ni on the recovery queue
3458                  * so we can continue the attempts to recover it.
3459                  */
3460                 if (ni->ni_recovery_state & LNET_NI_RECOVERY_FAILED) {
3461                         lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3462                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_FAILED;
3463                 }
3464
3465
3466                 lnet_ni_unlock(ni);
3467
3468                 if (now < ni->ni_next_ping) {
3469                         lnet_net_unlock(0);
3470                         continue;
3471                 }
3472
3473                 lnet_net_unlock(0);
3474
3475                 CDEBUG(D_NET, "attempting to recover local ni: %s\n",
3476                        libcfs_nidstr(&ni->ni_nid));
3477
3478                 lnet_ni_lock(ni);
3479                 if (!(ni->ni_recovery_state & LNET_NI_RECOVERY_PENDING)) {
3480                         ni->ni_recovery_state |= LNET_NI_RECOVERY_PENDING;
3481                         lnet_ni_unlock(ni);
3482
3483                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3484                         if (!ev_info) {
3485                                 CERROR("out of memory. Can't recover %s\n",
3486                                        libcfs_nidstr(&ni->ni_nid));
3487                                 lnet_ni_lock(ni);
3488                                 ni->ni_recovery_state &=
3489                                   ~LNET_NI_RECOVERY_PENDING;
3490                                 lnet_ni_unlock(ni);
3491                                 continue;
3492                         }
3493
3494                         mdh = ni->ni_ping_mdh;
3495                         /*
3496                          * Invalidate the ni mdh in case it's deleted.
3497                          * We'll unlink the mdh in this case below.
3498                          */
3499                         LNetInvalidateMDHandle(&ni->ni_ping_mdh);
3500                         /* FIXME need to handle large-addr nid */
3501                         nid = lnet_nid_to_nid4(&ni->ni_nid);
3502
3503                         /*
3504                          * remove the NI from the local queue and drop the
3505                          * reference count to it while we're recovering
3506                          * it. The reason for that, is that the NI could
3507                          * be deleted, and the way the code is structured
3508                          * is if we don't drop the NI, then the deletion
3509                          * code will enter a loop waiting for the
3510                          * reference count to be removed while holding the
3511                          * ln_mutex_lock(). When we look up the peer to
3512                          * send to in lnet_select_pathway() we will try to
3513                          * lock the ln_mutex_lock() as well, leading to
3514                          * a deadlock. By dropping the refcount and
3515                          * removing it from the list, we allow for the NI
3516                          * to be removed, then we use the cached NID to
3517                          * look it up again. If it's gone, then we just
3518                          * continue examining the rest of the queue.
3519                          */
3520                         lnet_net_lock(0);
3521                         list_del_init(&ni->ni_recovery);
3522                         lnet_ni_decref_locked(ni, 0);
3523                         lnet_net_unlock(0);
3524
3525                         ev_info->mt_type = MT_TYPE_LOCAL_NI;
3526                         ev_info->mt_nid = nid;
3527                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3528                                             ev_info, the_lnet.ln_mt_handler,
3529                                             true);
3530                         /* lookup the nid again */
3531                         lnet_net_lock(0);
3532                         ni = lnet_nid2ni_locked(nid, 0);
3533                         if (!ni) {
3534                                 /*
3535                                  * the NI has been deleted when we dropped
3536                                  * the ref count
3537                                  */
3538                                 lnet_net_unlock(0);
3539                                 LNetMDUnlink(mdh);
3540                                 continue;
3541                         }
3542                         ni->ni_ping_count++;
3543
3544                         ni->ni_ping_mdh = mdh;
3545                         lnet_ni_add_to_recoveryq_locked(ni, &processed_list,
3546                                                         now);
3547
3548                         if (rc) {
3549                                 lnet_ni_lock(ni);
3550                                 ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3551                                 lnet_ni_unlock(ni);
3552                         }
3553                         lnet_net_unlock(0);
3554                 } else
3555                         lnet_ni_unlock(ni);
3556         }
3557
3558         /*
3559          * put back the remaining NIs on the ln_mt_localNIRecovq to be
3560          * reexamined in the next iteration.
3561          */
3562         list_splice_init(&processed_list, &local_queue);
3563         lnet_net_lock(0);
3564         list_splice(&local_queue, &the_lnet.ln_mt_localNIRecovq);
3565         lnet_net_unlock(0);
3566 }
3567
3568 static int
3569 lnet_resendqs_create(void)
3570 {
3571         struct list_head **resendqs;
3572         resendqs = lnet_create_array_of_queues();
3573
3574         if (!resendqs)
3575                 return -ENOMEM;
3576
3577         lnet_net_lock(LNET_LOCK_EX);
3578         the_lnet.ln_mt_resendqs = resendqs;
3579         lnet_net_unlock(LNET_LOCK_EX);
3580
3581         return 0;
3582 }
3583
3584 static void
3585 lnet_clean_local_ni_recoveryq(void)
3586 {
3587         struct lnet_ni *ni;
3588
3589         /* This is only called when the monitor thread has stopped */
3590         lnet_net_lock(0);
3591
3592         while (!list_empty(&the_lnet.ln_mt_localNIRecovq)) {
3593                 ni = list_entry(the_lnet.ln_mt_localNIRecovq.next,
3594                                 struct lnet_ni, ni_recovery);
3595                 list_del_init(&ni->ni_recovery);
3596                 lnet_ni_lock(ni);
3597                 lnet_unlink_ni_recovery_mdh_locked(ni, 0, true);
3598                 lnet_ni_unlock(ni);
3599                 lnet_ni_decref_locked(ni, 0);
3600         }
3601
3602         lnet_net_unlock(0);
3603 }
3604
3605 static void
3606 lnet_unlink_lpni_recovery_mdh_locked(struct lnet_peer_ni *lpni, int cpt,
3607                                      bool force)
3608 {
3609         struct lnet_handle_md recovery_mdh;
3610
3611         LNetInvalidateMDHandle(&recovery_mdh);
3612
3613         if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING || force) {
3614                 recovery_mdh = lpni->lpni_recovery_ping_mdh;
3615                 LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3616         }
3617         spin_unlock(&lpni->lpni_lock);
3618         lnet_net_unlock(cpt);
3619         if (!LNetMDHandleIsInvalid(recovery_mdh))
3620                 LNetMDUnlink(recovery_mdh);
3621         lnet_net_lock(cpt);
3622         spin_lock(&lpni->lpni_lock);
3623 }
3624
3625 static void
3626 lnet_clean_peer_ni_recoveryq(void)
3627 {
3628         struct lnet_peer_ni *lpni, *tmp;
3629
3630         lnet_net_lock(LNET_LOCK_EX);
3631
3632         list_for_each_entry_safe(lpni, tmp, &the_lnet.ln_mt_peerNIRecovq,
3633                                  lpni_recovery) {
3634                 list_del_init(&lpni->lpni_recovery);
3635                 spin_lock(&lpni->lpni_lock);
3636                 lnet_unlink_lpni_recovery_mdh_locked(lpni, LNET_LOCK_EX, true);
3637                 spin_unlock(&lpni->lpni_lock);
3638                 lnet_peer_ni_decref_locked(lpni);
3639         }
3640
3641         lnet_net_unlock(LNET_LOCK_EX);
3642 }
3643
3644 static void
3645 lnet_clean_resendqs(void)
3646 {
3647         struct lnet_msg *msg, *tmp;
3648         LIST_HEAD(msgs);
3649         int i;
3650
3651         cfs_cpt_for_each(i, lnet_cpt_table()) {
3652                 lnet_net_lock(i);
3653                 list_splice_init(the_lnet.ln_mt_resendqs[i], &msgs);
3654                 lnet_net_unlock(i);
3655                 list_for_each_entry_safe(msg, tmp, &msgs, msg_list) {
3656                         list_del_init(&msg->msg_list);
3657                         msg->msg_no_resend = true;
3658                         lnet_finalize(msg, -ESHUTDOWN);
3659                 }
3660         }
3661
3662         cfs_percpt_free(the_lnet.ln_mt_resendqs);
3663 }
3664
3665 static void
3666 lnet_recover_peer_nis(void)
3667 {
3668         struct lnet_mt_event_info *ev_info;
3669         LIST_HEAD(processed_list);
3670         LIST_HEAD(local_queue);
3671         struct lnet_handle_md mdh;
3672         struct lnet_peer_ni *lpni;
3673         struct lnet_peer_ni *tmp;
3674         lnet_nid_t nid;
3675         int healthv;
3676         int rc;
3677         time64_t now;
3678
3679         /*
3680          * Always use cpt 0 for locking across all interactions with
3681          * ln_mt_peerNIRecovq
3682          */
3683         lnet_net_lock(0);
3684         list_splice_init(&the_lnet.ln_mt_peerNIRecovq,
3685                          &local_queue);
3686         lnet_net_unlock(0);
3687
3688         now = ktime_get_seconds();
3689
3690         list_for_each_entry_safe(lpni, tmp, &local_queue,
3691                                  lpni_recovery) {
3692                 /*
3693                  * The same protection strategy is used here as is in the
3694                  * local recovery case.
3695                  */
3696                 lnet_net_lock(0);
3697                 healthv = atomic_read(&lpni->lpni_healthv);
3698                 spin_lock(&lpni->lpni_lock);
3699                 if (lpni->lpni_state & LNET_PEER_NI_DELETING ||
3700                     healthv == LNET_MAX_HEALTH_VALUE) {
3701                         list_del_init(&lpni->lpni_recovery);
3702                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, false);
3703                         spin_unlock(&lpni->lpni_lock);
3704                         lnet_peer_ni_decref_locked(lpni);
3705                         lnet_net_unlock(0);
3706                         continue;
3707                 }
3708
3709                 /*
3710                  * If the peer NI has failed recovery we must unlink the
3711                  * md. But we want to keep the peer ni on the recovery
3712                  * queue so we can try to continue recovering it
3713                  */
3714                 if (lpni->lpni_state & LNET_PEER_NI_RECOVERY_FAILED) {
3715                         lnet_unlink_lpni_recovery_mdh_locked(lpni, 0, true);
3716                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_FAILED;
3717                 }
3718
3719                 spin_unlock(&lpni->lpni_lock);
3720
3721                 if (now < lpni->lpni_next_ping) {
3722                         lnet_net_unlock(0);
3723                         continue;
3724                 }
3725
3726                 lnet_net_unlock(0);
3727
3728                 /*
3729                  * NOTE: we're racing with peer deletion from user space.
3730                  * It's possible that a peer is deleted after we check its
3731                  * state. In this case the recovery can create a new peer
3732                  */
3733                 spin_lock(&lpni->lpni_lock);
3734                 if (!(lpni->lpni_state & LNET_PEER_NI_RECOVERY_PENDING) &&
3735                     !(lpni->lpni_state & LNET_PEER_NI_DELETING)) {
3736                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_PENDING;
3737                         spin_unlock(&lpni->lpni_lock);
3738
3739                         LIBCFS_ALLOC(ev_info, sizeof(*ev_info));
3740                         if (!ev_info) {
3741                                 CERROR("out of memory. Can't recover %s\n",
3742                                        libcfs_nidstr(&lpni->lpni_nid));
3743                                 spin_lock(&lpni->lpni_lock);
3744                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3745                                 spin_unlock(&lpni->lpni_lock);
3746                                 continue;
3747                         }
3748
3749                         /* look at the comments in lnet_recover_local_nis() */
3750                         mdh = lpni->lpni_recovery_ping_mdh;
3751                         LNetInvalidateMDHandle(&lpni->lpni_recovery_ping_mdh);
3752                         /* FIXME handle large-addr nid */
3753                         nid = lnet_nid_to_nid4(&lpni->lpni_nid);
3754                         lnet_net_lock(0);
3755                         list_del_init(&lpni->lpni_recovery);
3756                         lnet_peer_ni_decref_locked(lpni);
3757                         lnet_net_unlock(0);
3758
3759                         ev_info->mt_type = MT_TYPE_PEER_NI;
3760                         ev_info->mt_nid = nid;
3761                         rc = lnet_send_ping(nid, &mdh, LNET_INTERFACES_MIN,
3762                                             ev_info, the_lnet.ln_mt_handler,
3763                                             true);
3764                         lnet_net_lock(0);
3765                         /*
3766                          * lnet_find_peer_ni_locked() grabs a refcount for
3767                          * us. No need to take it explicitly.
3768                          */
3769                         lpni = lnet_find_peer_ni_locked(nid);
3770                         if (!lpni) {
3771                                 lnet_net_unlock(0);
3772                                 LNetMDUnlink(mdh);
3773                                 continue;
3774                         }
3775
3776                         lpni->lpni_ping_count++;
3777
3778                         lpni->lpni_recovery_ping_mdh = mdh;
3779
3780                         lnet_peer_ni_add_to_recoveryq_locked(lpni,
3781                                                              &processed_list,
3782                                                              now);
3783                         if (rc) {
3784                                 spin_lock(&lpni->lpni_lock);
3785                                 lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3786                                 spin_unlock(&lpni->lpni_lock);
3787                         }
3788
3789                         /* Drop the ref taken by lnet_find_peer_ni_locked() */
3790                         lnet_peer_ni_decref_locked(lpni);
3791                         lnet_net_unlock(0);
3792                 } else
3793                         spin_unlock(&lpni->lpni_lock);
3794         }
3795
3796         list_splice_init(&processed_list, &local_queue);
3797         lnet_net_lock(0);
3798         list_splice(&local_queue, &the_lnet.ln_mt_peerNIRecovq);
3799         lnet_net_unlock(0);
3800 }
3801
3802 static int
3803 lnet_monitor_thread(void *arg)
3804 {
3805         time64_t rsp_timeout = 0;
3806         time64_t now;
3807
3808         wait_for_completion(&the_lnet.ln_started);
3809         /*
3810          * The monitor thread takes care of the following:
3811          *  1. Checks the aliveness of routers
3812          *  2. Checks if there are messages on the resend queue to resend
3813          *     them.
3814          *  3. Check if there are any NIs on the local recovery queue and
3815          *     pings them
3816          *  4. Checks if there are any NIs on the remote recovery queue
3817          *     and pings them.
3818          */
3819         while (the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING) {
3820                 now = ktime_get_real_seconds();
3821
3822                 if (lnet_router_checker_active())
3823                         lnet_check_routers();
3824
3825                 lnet_resend_pending_msgs();
3826
3827                 if (now >= rsp_timeout) {
3828                         lnet_finalize_expired_responses();
3829                         rsp_timeout = now + (lnet_transaction_timeout / 2);
3830                 }
3831
3832                 lnet_recover_local_nis();
3833                 lnet_recover_peer_nis();
3834
3835                 /*
3836                  * TODO do we need to check if we should sleep without
3837                  * timeout?  Technically, an active system will always
3838                  * have messages in flight so this check will always
3839                  * evaluate to false. And on an idle system do we care
3840                  * if we wake up every 1 second? Although, we've seen
3841                  * cases where we get a complaint that an idle thread
3842                  * is waking up unnecessarily.
3843                  */
3844                 wait_for_completion_interruptible_timeout(
3845                         &the_lnet.ln_mt_wait_complete,
3846                         cfs_time_seconds(1));
3847                 /* Must re-init the completion before testing anything,
3848                  * including ln_mt_state.
3849                  */
3850                 reinit_completion(&the_lnet.ln_mt_wait_complete);
3851         }
3852
3853         /* Shutting down */
3854         lnet_net_lock(LNET_LOCK_EX);
3855         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
3856         lnet_net_unlock(LNET_LOCK_EX);
3857
3858         /* signal that the monitor thread is exiting */
3859         up(&the_lnet.ln_mt_signal);
3860
3861         return 0;
3862 }
3863
3864 /*
3865  * lnet_send_ping
3866  * Sends a ping.
3867  * Returns == 0 if success
3868  * Returns > 0 if LNetMDBind or prior fails
3869  * Returns < 0 if LNetGet fails
3870  */
3871 int
3872 lnet_send_ping(lnet_nid_t dest_nid,
3873                struct lnet_handle_md *mdh, int nnis,
3874                void *user_data, lnet_handler_t handler, bool recovery)
3875 {
3876         struct lnet_md md = { NULL };
3877         struct lnet_process_id id;
3878         struct lnet_ping_buffer *pbuf;
3879         int rc;
3880
3881         if (dest_nid == LNET_NID_ANY) {
3882                 rc = -EHOSTUNREACH;
3883                 goto fail_error;
3884         }
3885
3886         pbuf = lnet_ping_buffer_alloc(nnis, GFP_NOFS);
3887         if (!pbuf) {
3888                 rc = ENOMEM;
3889                 goto fail_error;
3890         }
3891
3892         /* initialize md content */
3893         md.start     = &pbuf->pb_info;
3894         md.length    = LNET_PING_INFO_SIZE(nnis);
3895         md.threshold = 2; /* GET/REPLY */
3896         md.max_size  = 0;
3897         md.options   = LNET_MD_TRUNCATE | LNET_MD_TRACK_RESPONSE;
3898         md.user_ptr  = user_data;
3899         md.handler   = handler;
3900
3901         rc = LNetMDBind(&md, LNET_UNLINK, mdh);
3902         if (rc) {
3903                 lnet_ping_buffer_decref(pbuf);
3904                 CERROR("Can't bind MD: %d\n", rc);
3905                 rc = -rc; /* change the rc to positive */
3906                 goto fail_error;
3907         }
3908         id.pid = LNET_PID_LUSTRE;
3909         id.nid = dest_nid;
3910
3911         rc = LNetGet(LNET_NID_ANY, *mdh, id,
3912                      LNET_RESERVED_PORTAL,
3913                      LNET_PROTO_PING_MATCHBITS, 0, recovery);
3914
3915         if (rc)
3916                 goto fail_unlink_md;
3917
3918         return 0;
3919
3920 fail_unlink_md:
3921         LNetMDUnlink(*mdh);
3922         LNetInvalidateMDHandle(mdh);
3923 fail_error:
3924         return rc;
3925 }
3926
3927 static void
3928 lnet_handle_recovery_reply(struct lnet_mt_event_info *ev_info,
3929                            int status, bool send, bool unlink_event)
3930 {
3931         lnet_nid_t nid = ev_info->mt_nid;
3932
3933         if (ev_info->mt_type == MT_TYPE_LOCAL_NI) {
3934                 struct lnet_ni *ni;
3935
3936                 lnet_net_lock(0);
3937                 ni = lnet_nid2ni_locked(nid, 0);
3938                 if (!ni) {
3939                         lnet_net_unlock(0);
3940                         return;
3941                 }
3942                 lnet_ni_lock(ni);
3943                 if (!send || (send && status != 0))
3944                         ni->ni_recovery_state &= ~LNET_NI_RECOVERY_PENDING;
3945                 if (status)
3946                         ni->ni_recovery_state |= LNET_NI_RECOVERY_FAILED;
3947                 lnet_ni_unlock(ni);
3948                 lnet_net_unlock(0);
3949
3950                 if (status != 0) {
3951                         CERROR("local NI (%s) recovery failed with %d\n",
3952                                libcfs_nid2str(nid), status);
3953                         return;
3954                 }
3955                 /*
3956                  * need to increment healthv for the ni here, because in
3957                  * the lnet_finalize() path we don't have access to this
3958                  * NI. And in order to get access to it, we'll need to
3959                  * carry forward too much information.
3960                  * In the peer case, it'll naturally be incremented
3961                  */
3962                 if (!unlink_event)
3963                         lnet_inc_healthv(&ni->ni_healthv,
3964                                          lnet_health_sensitivity);
3965         } else {
3966                 struct lnet_peer_ni *lpni;
3967                 int cpt;
3968
3969                 cpt = lnet_net_lock_current();
3970                 lpni = lnet_find_peer_ni_locked(nid);
3971                 if (!lpni) {
3972                         lnet_net_unlock(cpt);
3973                         return;
3974                 }
3975                 spin_lock(&lpni->lpni_lock);
3976                 if (!send || (send && status != 0))
3977                         lpni->lpni_state &= ~LNET_PEER_NI_RECOVERY_PENDING;
3978                 if (status)
3979                         lpni->lpni_state |= LNET_PEER_NI_RECOVERY_FAILED;
3980                 spin_unlock(&lpni->lpni_lock);
3981                 lnet_peer_ni_decref_locked(lpni);
3982                 lnet_net_unlock(cpt);
3983
3984                 if (status != 0)
3985                         CERROR("peer NI (%s) recovery failed with %d\n",
3986                                libcfs_nid2str(nid), status);
3987         }
3988 }
3989
3990 void
3991 lnet_mt_event_handler(struct lnet_event *event)
3992 {
3993         struct lnet_mt_event_info *ev_info = event->md_user_ptr;
3994         struct lnet_ping_buffer *pbuf;
3995
3996         /* TODO: remove assert */
3997         LASSERT(event->type == LNET_EVENT_REPLY ||
3998                 event->type == LNET_EVENT_SEND ||
3999                 event->type == LNET_EVENT_UNLINK);
4000
4001         CDEBUG(D_NET, "Received event: %d status: %d\n", event->type,
4002                event->status);
4003
4004         switch (event->type) {
4005         case LNET_EVENT_UNLINK:
4006                 CDEBUG(D_NET, "%s recovery ping unlinked\n",
4007                        libcfs_nid2str(ev_info->mt_nid));
4008                 /* fallthrough */
4009         case LNET_EVENT_REPLY:
4010                 lnet_handle_recovery_reply(ev_info, event->status, false,
4011                                            event->type == LNET_EVENT_UNLINK);
4012                 break;
4013         case LNET_EVENT_SEND:
4014                 CDEBUG(D_NET, "%s recovery message sent %s:%d\n",
4015                                libcfs_nid2str(ev_info->mt_nid),
4016                                (event->status) ? "unsuccessfully" :
4017                                "successfully", event->status);
4018                 lnet_handle_recovery_reply(ev_info, event->status, true, false);
4019                 break;
4020         default:
4021                 CERROR("Unexpected event: %d\n", event->type);
4022                 break;
4023         }
4024         if (event->unlinked) {
4025                 LIBCFS_FREE(ev_info, sizeof(*ev_info));
4026                 pbuf = LNET_PING_INFO_TO_BUFFER(event->md_start);
4027                 lnet_ping_buffer_decref(pbuf);
4028         }
4029 }
4030
4031 static int
4032 lnet_rsp_tracker_create(void)
4033 {
4034         struct list_head **rstqs;
4035         rstqs = lnet_create_array_of_queues();
4036
4037         if (!rstqs)
4038                 return -ENOMEM;
4039
4040         the_lnet.ln_mt_rstq = rstqs;
4041
4042         return 0;
4043 }
4044
4045 static void
4046 lnet_rsp_tracker_clean(void)
4047 {
4048         lnet_finalize_expired_responses();
4049
4050         cfs_percpt_free(the_lnet.ln_mt_rstq);
4051         the_lnet.ln_mt_rstq = NULL;
4052 }
4053
4054 int lnet_monitor_thr_start(void)
4055 {
4056         int rc = 0;
4057         struct task_struct *task;
4058
4059         if (the_lnet.ln_mt_state != LNET_MT_STATE_SHUTDOWN)
4060                 return -EALREADY;
4061
4062         rc = lnet_resendqs_create();
4063         if (rc)
4064                 return rc;
4065
4066         rc = lnet_rsp_tracker_create();
4067         if (rc)
4068                 goto clean_queues;
4069
4070         sema_init(&the_lnet.ln_mt_signal, 0);
4071
4072         lnet_net_lock(LNET_LOCK_EX);
4073         the_lnet.ln_mt_state = LNET_MT_STATE_RUNNING;
4074         lnet_net_unlock(LNET_LOCK_EX);
4075         task = kthread_run(lnet_monitor_thread, NULL, "monitor_thread");
4076         if (IS_ERR(task)) {
4077                 rc = PTR_ERR(task);
4078                 CERROR("Can't start monitor thread: %d\n", rc);
4079                 goto clean_thread;
4080         }
4081
4082         return 0;
4083
4084 clean_thread:
4085         lnet_net_lock(LNET_LOCK_EX);
4086         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4087         lnet_net_unlock(LNET_LOCK_EX);
4088         /* block until event callback signals exit */
4089         down(&the_lnet.ln_mt_signal);
4090         /* clean up */
4091         lnet_net_lock(LNET_LOCK_EX);
4092         the_lnet.ln_mt_state = LNET_MT_STATE_SHUTDOWN;
4093         lnet_net_unlock(LNET_LOCK_EX);
4094         lnet_rsp_tracker_clean();
4095         lnet_clean_local_ni_recoveryq();
4096         lnet_clean_peer_ni_recoveryq();
4097         lnet_clean_resendqs();
4098         the_lnet.ln_mt_handler = NULL;
4099         return rc;
4100 clean_queues:
4101         lnet_rsp_tracker_clean();
4102         lnet_clean_local_ni_recoveryq();
4103         lnet_clean_peer_ni_recoveryq();
4104         lnet_clean_resendqs();
4105         return rc;
4106 }
4107
4108 void lnet_monitor_thr_stop(void)
4109 {
4110         if (the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN)
4111                 return;
4112
4113         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_RUNNING);
4114         lnet_net_lock(LNET_LOCK_EX);
4115         the_lnet.ln_mt_state = LNET_MT_STATE_STOPPING;
4116         lnet_net_unlock(LNET_LOCK_EX);
4117
4118         /* tell the monitor thread that we're shutting down */
4119         complete(&the_lnet.ln_mt_wait_complete);
4120
4121         /* block until monitor thread signals that it's done */
4122         down(&the_lnet.ln_mt_signal);
4123         LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
4124
4125         /* perform cleanup tasks */
4126         lnet_rsp_tracker_clean();
4127         lnet_clean_local_ni_recoveryq();
4128         lnet_clean_peer_ni_recoveryq();
4129         lnet_clean_resendqs();
4130 }
4131
4132 void
4133 lnet_drop_message(struct lnet_ni *ni, int cpt, void *private, unsigned int nob,
4134                   __u32 msg_type)
4135 {
4136         lnet_net_lock(cpt);
4137         lnet_incr_stats(&ni->ni_stats, msg_type, LNET_STATS_TYPE_DROP);
4138         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
4139         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length += nob;
4140         lnet_net_unlock(cpt);
4141
4142         lnet_ni_recv(ni, private, NULL, 0, 0, 0, nob);
4143 }
4144
4145 static void
4146 lnet_recv_put(struct lnet_ni *ni, struct lnet_msg *msg)
4147 {
4148         struct lnet_hdr *hdr = &msg->msg_hdr;
4149
4150         if (msg->msg_wanted != 0)
4151                 lnet_setpayloadbuffer(msg);
4152
4153         lnet_build_msg_event(msg, LNET_EVENT_PUT);
4154
4155         /* Must I ACK?  If so I'll grab the ack_wmd out of the header and put
4156          * it back into the ACK during lnet_finalize() */
4157         msg->msg_ack = (!lnet_is_wire_handle_none(&hdr->msg.put.ack_wmd) &&
4158                         (msg->msg_md->md_options & LNET_MD_ACK_DISABLE) == 0);
4159
4160         lnet_ni_recv(ni, msg->msg_private, msg, msg->msg_rx_delayed,
4161                      msg->msg_offset, msg->msg_wanted, hdr->payload_length);
4162 }
4163
4164 static int
4165 lnet_parse_put(struct lnet_ni *ni, struct lnet_msg *msg)
4166 {
4167         struct lnet_hdr         *hdr = &msg->msg_hdr;
4168         struct lnet_match_info  info;
4169         int                     rc;
4170         bool                    ready_delay;
4171
4172         /* Convert put fields to host byte order */
4173         hdr->msg.put.match_bits = le64_to_cpu(hdr->msg.put.match_bits);
4174         hdr->msg.put.ptl_index  = le32_to_cpu(hdr->msg.put.ptl_index);
4175         hdr->msg.put.offset     = le32_to_cpu(hdr->msg.put.offset);
4176
4177         /* Primary peer NID. */
4178         lnet_nid4_to_nid(msg->msg_initiator, &info.mi_id.nid);
4179         info.mi_id.pid  = hdr->src_pid;
4180         info.mi_opc     = LNET_MD_OP_PUT;
4181         info.mi_portal  = hdr->msg.put.ptl_index;
4182         info.mi_rlength = hdr->payload_length;
4183         info.mi_roffset = hdr->msg.put.offset;
4184         info.mi_mbits   = hdr->msg.put.match_bits;
4185         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
4186
4187         msg->msg_rx_ready_delay = ni->ni_net->net_lnd->lnd_eager_recv == NULL;
4188         ready_delay = msg->msg_rx_ready_delay;
4189
4190  again:
4191         rc = lnet_ptl_match_md(&info, msg);
4192         switch (rc) {
4193         default:
4194                 LBUG();
4195
4196         case LNET_MATCHMD_OK:
4197                 lnet_recv_put(ni, msg);
4198                 return 0;
4199
4200         case LNET_MATCHMD_NONE:
4201                 if (ready_delay)
4202                         /* no eager_recv or has already called it, should
4203                          * have been attached on delayed list */
4204                         return 0;
4205
4206                 rc = lnet_ni_eager_recv(ni, msg);
4207                 if (rc == 0) {
4208                         ready_delay = true;
4209                         goto again;
4210                 }
4211                 /* fall through */
4212
4213         case LNET_MATCHMD_DROP:
4214                 CNETERR("Dropping PUT from %s portal %d match %llu"
4215                         " offset %d length %d: %d\n",
4216                         libcfs_idstr(&info.mi_id), info.mi_portal,
4217                         info.mi_mbits, info.mi_roffset, info.mi_rlength, rc);
4218
4219                 return -ENOENT; /* -ve: OK but no match */
4220         }
4221 }
4222
4223 static int
4224 lnet_parse_get(struct lnet_ni *ni, struct lnet_msg *msg, int rdma_get)
4225 {
4226         struct lnet_match_info info;
4227         struct lnet_hdr *hdr = &msg->msg_hdr;
4228         struct lnet_process_id source_id;
4229         struct lnet_handle_wire reply_wmd;
4230         int rc;
4231
4232         /* Convert get fields to host byte order */
4233         hdr->msg.get.match_bits   = le64_to_cpu(hdr->msg.get.match_bits);
4234         hdr->msg.get.ptl_index    = le32_to_cpu(hdr->msg.get.ptl_index);
4235         hdr->msg.get.sink_length  = le32_to_cpu(hdr->msg.get.sink_length);
4236         hdr->msg.get.src_offset   = le32_to_cpu(hdr->msg.get.src_offset);
4237
4238         source_id.nid = hdr->src_nid;
4239         source_id.pid = hdr->src_pid;
4240         /* Primary peer NID */
4241         lnet_nid4_to_nid(msg->msg_initiator, &info.mi_id.nid);
4242         info.mi_id.pid  = hdr->src_pid;
4243         info.mi_opc     = LNET_MD_OP_GET;
4244         info.mi_portal  = hdr->msg.get.ptl_index;
4245         info.mi_rlength = hdr->msg.get.sink_length;
4246         info.mi_roffset = hdr->msg.get.src_offset;
4247         info.mi_mbits   = hdr->msg.get.match_bits;
4248         info.mi_cpt     = lnet_cpt_of_nid(msg->msg_initiator, ni);
4249
4250         rc = lnet_ptl_match_md(&info, msg);
4251         if (rc == LNET_MATCHMD_DROP) {
4252                 CNETERR("Dropping GET from %s portal %d match %llu"
4253                         " offset %d length %d\n",
4254                         libcfs_idstr(&info.mi_id), info.mi_portal,
4255                         info.mi_mbits, info.mi_roffset, info.mi_rlength);
4256                 return -ENOENT; /* -ve: OK but no match */
4257         }
4258
4259         LASSERT(rc == LNET_MATCHMD_OK);
4260
4261         lnet_build_msg_event(msg, LNET_EVENT_GET);
4262
4263         reply_wmd = hdr->msg.get.return_wmd;
4264
4265         lnet_prep_send(msg, LNET_MSG_REPLY, source_id,
4266                        msg->msg_offset, msg->msg_wanted);
4267
4268         msg->msg_hdr.msg.reply.dst_wmd = reply_wmd;
4269
4270         if (rdma_get) {
4271                 /* The LND completes the REPLY from her recv procedure */
4272                 lnet_ni_recv(ni, msg->msg_private, msg, 0,
4273                              msg->msg_offset, msg->msg_len, msg->msg_len);
4274                 return 0;
4275         }
4276
4277         lnet_ni_recv(ni, msg->msg_private, NULL, 0, 0, 0, 0);
4278         msg->msg_receiving = 0;
4279
4280         /* FIXME need to handle large-addr nid */
4281         rc = lnet_send(lnet_nid_to_nid4(&ni->ni_nid), msg, msg->msg_from);
4282         if (rc < 0) {
4283                 /* didn't get as far as lnet_ni_send() */
4284                 CERROR("%s: Unable to send REPLY for GET from %s: %d\n",
4285                        libcfs_nidstr(&ni->ni_nid),
4286                        libcfs_idstr(&info.mi_id), rc);
4287
4288                 lnet_finalize(msg, rc);
4289         }
4290
4291         return 0;
4292 }
4293
4294 static int
4295 lnet_parse_reply(struct lnet_ni *ni, struct lnet_msg *msg)
4296 {
4297         void *private = msg->msg_private;
4298         struct lnet_hdr *hdr = &msg->msg_hdr;
4299         struct lnet_process_id src = {0};
4300         struct lnet_libmd *md;
4301         unsigned int rlength;
4302         unsigned int mlength;
4303         int cpt;
4304
4305         cpt = lnet_cpt_of_cookie(hdr->msg.reply.dst_wmd.wh_object_cookie);
4306         lnet_res_lock(cpt);
4307
4308         src.nid = hdr->src_nid;
4309         src.pid = hdr->src_pid;
4310
4311         /* NB handles only looked up by creator (no flips) */
4312         md = lnet_wire_handle2md(&hdr->msg.reply.dst_wmd);
4313         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4314                 CNETERR("%s: Dropping REPLY from %s for %s "
4315                         "MD %#llx.%#llx\n",
4316                         libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4317                         (md == NULL) ? "invalid" : "inactive",
4318                         hdr->msg.reply.dst_wmd.wh_interface_cookie,
4319                         hdr->msg.reply.dst_wmd.wh_object_cookie);
4320                 if (md != NULL && md->md_me != NULL)
4321                         CERROR("REPLY MD also attached to portal %d\n",
4322                                md->md_me->me_portal);
4323
4324                 lnet_res_unlock(cpt);
4325                 return -ENOENT; /* -ve: OK but no match */
4326         }
4327
4328         LASSERT(md->md_offset == 0);
4329
4330         rlength = hdr->payload_length;
4331         mlength = min(rlength, md->md_length);
4332
4333         if (mlength < rlength &&
4334             (md->md_options & LNET_MD_TRUNCATE) == 0) {
4335                 CNETERR("%s: Dropping REPLY from %s length %d "
4336                         "for MD %#llx would overflow (%d)\n",
4337                         libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4338                         rlength, hdr->msg.reply.dst_wmd.wh_object_cookie,
4339                         mlength);
4340                 lnet_res_unlock(cpt);
4341                 return -ENOENT; /* -ve: OK but no match */
4342         }
4343
4344         CDEBUG(D_NET, "%s: Reply from %s of length %d/%d into md %#llx\n",
4345                libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4346                mlength, rlength, hdr->msg.reply.dst_wmd.wh_object_cookie);
4347
4348         lnet_msg_attach_md(msg, md, 0, mlength);
4349
4350         if (mlength != 0)
4351                 lnet_setpayloadbuffer(msg);
4352
4353         lnet_res_unlock(cpt);
4354
4355         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
4356
4357         lnet_ni_recv(ni, private, msg, 0, 0, mlength, rlength);
4358         return 0;
4359 }
4360
4361 static int
4362 lnet_parse_ack(struct lnet_ni *ni, struct lnet_msg *msg)
4363 {
4364         struct lnet_hdr *hdr = &msg->msg_hdr;
4365         struct lnet_process_id src = {0};
4366         struct lnet_libmd *md;
4367         int cpt;
4368
4369         src.nid = hdr->src_nid;
4370         src.pid = hdr->src_pid;
4371
4372         /* Convert ack fields to host byte order */
4373         hdr->msg.ack.match_bits = le64_to_cpu(hdr->msg.ack.match_bits);
4374         hdr->msg.ack.mlength = le32_to_cpu(hdr->msg.ack.mlength);
4375
4376         cpt = lnet_cpt_of_cookie(hdr->msg.ack.dst_wmd.wh_object_cookie);
4377         lnet_res_lock(cpt);
4378
4379         /* NB handles only looked up by creator (no flips) */
4380         md = lnet_wire_handle2md(&hdr->msg.ack.dst_wmd);
4381         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
4382                 /* Don't moan; this is expected */
4383                 CDEBUG(D_NET,
4384                        "%s: Dropping ACK from %s to %s MD %#llx.%#llx\n",
4385                        libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4386                        (md == NULL) ? "invalid" : "inactive",
4387                        hdr->msg.ack.dst_wmd.wh_interface_cookie,
4388                        hdr->msg.ack.dst_wmd.wh_object_cookie);
4389                 if (md != NULL && md->md_me != NULL)
4390                         CERROR("Source MD also attached to portal %d\n",
4391                                md->md_me->me_portal);
4392
4393                 lnet_res_unlock(cpt);
4394                 return -ENOENT;                  /* -ve! */
4395         }
4396
4397         CDEBUG(D_NET, "%s: ACK from %s into md %#llx\n",
4398                libcfs_nidstr(&ni->ni_nid), libcfs_id2str(src),
4399                hdr->msg.ack.dst_wmd.wh_object_cookie);
4400
4401         lnet_msg_attach_md(msg, md, 0, 0);
4402
4403         lnet_res_unlock(cpt);
4404
4405         lnet_build_msg_event(msg, LNET_EVENT_ACK);
4406
4407         lnet_ni_recv(ni, msg->msg_private, msg, 0, 0, 0, msg->msg_len);
4408         return 0;
4409 }
4410
4411 /**
4412  * \retval LNET_CREDIT_OK       If \a msg is forwarded
4413  * \retval LNET_CREDIT_WAIT     If \a msg is blocked because w/o buffer
4414  * \retval -ve                  error code
4415  */
4416 int
4417 lnet_parse_forward_locked(struct lnet_ni *ni, struct lnet_msg *msg)
4418 {
4419         int     rc = 0;
4420
4421         if (!the_lnet.ln_routing)
4422                 return -ECANCELED;
4423
4424         if (msg->msg_rxpeer->lpni_rtrcredits <= 0 ||
4425             lnet_msg2bufpool(msg)->rbp_credits <= 0) {
4426                 if (ni->ni_net->net_lnd->lnd_eager_recv == NULL) {
4427                         msg->msg_rx_ready_delay = 1;
4428                 } else {
4429                         lnet_net_unlock(msg->msg_rx_cpt);
4430                         rc = lnet_ni_eager_recv(ni, msg);
4431                         lnet_net_lock(msg->msg_rx_cpt);
4432                 }
4433         }
4434
4435         if (rc == 0)
4436                 rc = lnet_post_routed_recv_locked(msg, 0);
4437         return rc;
4438 }
4439
4440 int
4441 lnet_parse_local(struct lnet_ni *ni, struct lnet_msg *msg)
4442 {
4443         int     rc;
4444
4445         switch (msg->msg_type) {
4446         case LNET_MSG_ACK:
4447                 rc = lnet_parse_ack(ni, msg);
4448                 break;
4449         case LNET_MSG_PUT:
4450                 rc = lnet_parse_put(ni, msg);
4451                 break;
4452         case LNET_MSG_GET:
4453                 rc = lnet_parse_get(ni, msg, msg->msg_rdma_get);
4454                 break;
4455         case LNET_MSG_REPLY:
4456                 rc = lnet_parse_reply(ni, msg);
4457                 break;
4458         default: /* prevent an unused label if !kernel */
4459                 LASSERT(0);
4460                 return -EPROTO;
4461         }
4462
4463         LASSERT(rc == 0 || rc == -ENOENT);
4464         return rc;
4465 }
4466
4467 char *
4468 lnet_msgtyp2str (int type)
4469 {
4470         switch (type) {
4471         case LNET_MSG_ACK:
4472                 return ("ACK");
4473         case LNET_MSG_PUT:
4474                 return ("PUT");
4475         case LNET_MSG_GET:
4476                 return ("GET");
4477         case LNET_MSG_REPLY:
4478                 return ("REPLY");
4479         case LNET_MSG_HELLO:
4480                 return ("HELLO");
4481         default:
4482                 return ("<UNKNOWN>");
4483         }
4484 }
4485
4486 int
4487 lnet_parse(struct lnet_ni *ni, struct lnet_hdr *hdr, lnet_nid_t from_nid4,
4488            void *private, int rdma_req)
4489 {
4490         struct lnet_peer_ni *lpni;
4491         struct lnet_msg *msg;
4492         __u32 payload_length;
4493         lnet_pid_t dest_pid;
4494         lnet_nid_t dest_nid;
4495         lnet_nid_t src_nid;
4496         struct lnet_nid from_nid;
4497         bool push = false;
4498         int for_me;
4499         __u32 type;
4500         int rc = 0;
4501         int cpt;
4502
4503         LASSERT (!in_interrupt ());
4504
4505         lnet_nid4_to_nid(from_nid4, &from_nid);
4506
4507         type = le32_to_cpu(hdr->type);
4508         src_nid = le64_to_cpu(hdr->src_nid);
4509         dest_nid = le64_to_cpu(hdr->dest_nid);
4510         dest_pid = le32_to_cpu(hdr->dest_pid);
4511         payload_length = le32_to_cpu(hdr->payload_length);
4512
4513         /* FIXME handle large-addr nids */
4514         for_me = (lnet_nid_to_nid4(&ni->ni_nid) == dest_nid);
4515         cpt = lnet_cpt_of_nid(from_nid4, ni);
4516
4517         CDEBUG(D_NET, "TRACE: %s(%s) <- %s : %s - %s\n",
4518                 libcfs_nid2str(dest_nid),
4519                 libcfs_nidstr(&ni->ni_nid),
4520                 libcfs_nid2str(src_nid),
4521                 lnet_msgtyp2str(type),
4522                 (for_me) ? "for me" : "routed");
4523
4524         switch (type) {
4525         case LNET_MSG_ACK:
4526         case LNET_MSG_GET:
4527                 if (payload_length > 0) {
4528                         CERROR("%s, src %s: bad %s payload %d (0 expected)\n",
4529                                libcfs_nid2str(from_nid4),
4530                                libcfs_nid2str(src_nid),
4531                                lnet_msgtyp2str(type), payload_length);
4532                         return -EPROTO;
4533                 }
4534                 break;
4535
4536         case LNET_MSG_PUT:
4537         case LNET_MSG_REPLY:
4538                 if (payload_length >
4539                     (__u32)(for_me ? LNET_MAX_PAYLOAD : LNET_MTU)) {
4540                         CERROR("%s, src %s: bad %s payload %d "
4541                                "(%d max expected)\n",
4542                                libcfs_nid2str(from_nid4),
4543                                libcfs_nid2str(src_nid),
4544                                lnet_msgtyp2str(type),
4545                                payload_length,
4546                                for_me ? LNET_MAX_PAYLOAD : LNET_MTU);
4547                         return -EPROTO;
4548                 }
4549                 break;
4550
4551         default:
4552                 CERROR("%s, src %s: Bad message type 0x%x\n",
4553                        libcfs_nid2str(from_nid4),
4554                        libcfs_nid2str(src_nid), type);
4555                 return -EPROTO;
4556         }
4557
4558         if (the_lnet.ln_routing &&
4559             ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
4560                 lnet_ni_lock(ni);
4561                 spin_lock(&ni->ni_net->net_lock);
4562                 ni->ni_net->net_last_alive = ktime_get_real_seconds();
4563                 spin_unlock(&ni->ni_net->net_lock);
4564                 push = lnet_ni_set_status_locked(ni, LNET_NI_STATUS_UP);
4565                 lnet_ni_unlock(ni);
4566         }
4567
4568         if (push)
4569                 lnet_push_update_to_peers(1);
4570
4571         /* Regard a bad destination NID as a protocol error.  Senders should
4572          * know what they're doing; if they don't they're misconfigured, buggy
4573          * or malicious so we chop them off at the knees :) */
4574
4575         if (!for_me) {
4576                 if (LNET_NIDNET(dest_nid) == LNET_NID_NET(&ni->ni_nid)) {
4577                         /* should have gone direct */
4578                         CERROR("%s, src %s: Bad dest nid %s "
4579                                "(should have been sent direct)\n",
4580                                 libcfs_nid2str(from_nid4),
4581                                 libcfs_nid2str(src_nid),
4582                                 libcfs_nid2str(dest_nid));
4583                         return -EPROTO;
4584                 }
4585
4586                 if (lnet_islocalnid4(dest_nid)) {
4587                         /* dest is another local NI; sender should have used
4588                          * this node's NID on its own network */
4589                         CERROR("%s, src %s: Bad dest nid %s "
4590                                "(it's my nid but on a different network)\n",
4591                                 libcfs_nid2str(from_nid4),
4592                                 libcfs_nid2str(src_nid),
4593                                 libcfs_nid2str(dest_nid));
4594                         return -EPROTO;
4595                 }
4596
4597                 if (rdma_req && type == LNET_MSG_GET) {
4598                         CERROR("%s, src %s: Bad optimized GET for %s "
4599                                "(final destination must be me)\n",
4600                                 libcfs_nid2str(from_nid4),
4601                                 libcfs_nid2str(src_nid),
4602                                 libcfs_nid2str(dest_nid));
4603                         return -EPROTO;
4604                 }
4605
4606                 if (!the_lnet.ln_routing) {
4607                         CERROR("%s, src %s: Dropping message for %s "
4608                                "(routing not enabled)\n",
4609                                 libcfs_nid2str(from_nid4),
4610                                 libcfs_nid2str(src_nid),
4611                                 libcfs_nid2str(dest_nid));
4612                         goto drop;
4613                 }
4614         }
4615
4616         /* Message looks OK; we're not going to return an error, so we MUST
4617          * call back lnd_recv() come what may... */
4618
4619         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4620             fail_peer(src_nid, 0)) {                    /* shall we now? */
4621                 CERROR("%s, src %s: Dropping %s to simulate failure\n",
4622                        libcfs_nid2str(from_nid4), libcfs_nid2str(src_nid),
4623                        lnet_msgtyp2str(type));
4624                 goto drop;
4625         }
4626
4627         /* FIXME need to support large-addr nid */
4628         if (!list_empty(&the_lnet.ln_drop_rules) &&
4629             lnet_drop_rule_match(hdr, lnet_nid_to_nid4(&ni->ni_nid), NULL)) {
4630                 CDEBUG(D_NET,
4631                        "%s, src %s, dst %s: Dropping %s to simulate silent message loss\n",
4632                        libcfs_nid2str(from_nid4), libcfs_nid2str(src_nid),
4633                        libcfs_nid2str(dest_nid), lnet_msgtyp2str(type));
4634                 goto drop;
4635         }
4636
4637         msg = lnet_msg_alloc();
4638         if (msg == NULL) {
4639                 CERROR("%s, src %s: Dropping %s (out of memory)\n",
4640                        libcfs_nid2str(from_nid4), libcfs_nid2str(src_nid),
4641                        lnet_msgtyp2str(type));
4642                 goto drop;
4643         }
4644
4645         /* msg zeroed in lnet_msg_alloc; i.e. flags all clear,
4646          * pointers NULL etc */
4647
4648         msg->msg_type = type;
4649         msg->msg_private = private;
4650         msg->msg_receiving = 1;
4651         msg->msg_rdma_get = rdma_req;
4652         msg->msg_len = msg->msg_wanted = payload_length;
4653         msg->msg_offset = 0;
4654         msg->msg_hdr = *hdr;
4655         /* for building message event */
4656         msg->msg_from = from_nid4;
4657         if (!for_me) {
4658                 msg->msg_target.pid     = dest_pid;
4659                 msg->msg_target.nid     = dest_nid;
4660                 msg->msg_routing        = 1;
4661
4662         } else {
4663                 /* convert common msg->hdr fields to host byteorder */
4664                 msg->msg_hdr.type       = type;
4665                 msg->msg_hdr.src_nid    = src_nid;
4666                 msg->msg_hdr.src_pid    = le32_to_cpu(msg->msg_hdr.src_pid);
4667                 msg->msg_hdr.dest_nid   = dest_nid;
4668                 msg->msg_hdr.dest_pid   = dest_pid;
4669                 msg->msg_hdr.payload_length = payload_length;
4670         }
4671
4672         lnet_net_lock(cpt);
4673         lpni = lnet_peerni_by_nid_locked(&from_nid, &ni->ni_nid, cpt);
4674         if (IS_ERR(lpni)) {
4675                 lnet_net_unlock(cpt);
4676                 rc = PTR_ERR(lpni);
4677                 CERROR("%s, src %s: Dropping %s (error %d looking up sender)\n",
4678                        libcfs_nid2str(from_nid4), libcfs_nid2str(src_nid),
4679                        lnet_msgtyp2str(type), rc);
4680                 lnet_msg_free(msg);
4681                 if (rc == -ESHUTDOWN)
4682                         /* We are shutting down.  Don't do anything more */
4683                         return 0;
4684                 goto drop;
4685         }
4686
4687         /* If this message was forwarded to us from a router then we may need
4688          * to update router aliveness or check for an asymmetrical route
4689          * (or both)
4690          */
4691         if (((lnet_drop_asym_route && for_me) ||
4692              !lpni->lpni_peer_net->lpn_peer->lp_alive) &&
4693             LNET_NIDNET(src_nid) != LNET_NIDNET(from_nid4)) {
4694                 __u32 src_net_id = LNET_NIDNET(src_nid);
4695                 struct lnet_peer *gw = lpni->lpni_peer_net->lpn_peer;
4696                 struct lnet_route *route;
4697                 bool found = false;
4698
4699                 list_for_each_entry(route, &gw->lp_routes, lr_gwlist) {
4700                         if (route->lr_net == src_net_id) {
4701                                 found = true;
4702                                 /* If we're transitioning the gateway from
4703                                  * dead -> alive, and discovery is disabled
4704                                  * locally or on the gateway, then we need to
4705                                  * update the cached route aliveness for each
4706                                  * route to the src_nid's net.
4707                                  *
4708                                  * Otherwise, we're only checking for
4709                                  * symmetrical route, and we can break the
4710                                  * loop
4711                                  */
4712                                 if (!gw->lp_alive &&
4713                                     lnet_is_discovery_disabled(gw))
4714                                         lnet_set_route_aliveness(route, true);
4715                                 else
4716                                         break;
4717                         }
4718                 }
4719                 if (lnet_drop_asym_route && for_me && !found) {
4720                         /* Drop ref taken by lnet_nid2peerni_locked() */
4721                         lnet_peer_ni_decref_locked(lpni);
4722                         lnet_net_unlock(cpt);
4723                         /* we would not use from_nid to route a message to
4724                          * src_nid
4725                          * => asymmetric routing detected but forbidden
4726                          */
4727                         CERROR("%s, src %s: Dropping asymmetrical route %s\n",
4728                                libcfs_nid2str(from_nid4),
4729                                libcfs_nid2str(src_nid), lnet_msgtyp2str(type));
4730                         lnet_msg_free(msg);
4731                         goto drop;
4732                 }
4733                 if (!gw->lp_alive) {
4734                         struct lnet_peer_net *lpn;
4735                         struct lnet_peer_ni *lpni2;
4736
4737                         gw->lp_alive = true;
4738                         /* Mark all remote NIs on src_nid's net UP */
4739                         lpn = lnet_peer_get_net_locked(gw, src_net_id);
4740                         if (lpn)
4741                                 list_for_each_entry(lpni2, &lpn->lpn_peer_nis,
4742                                                     lpni_peer_nis)
4743                                         lpni2->lpni_ns_status = LNET_NI_STATUS_UP;
4744                 }
4745         }
4746
4747         lpni->lpni_last_alive = ktime_get_seconds();
4748
4749         msg->msg_rxpeer = lpni;
4750         msg->msg_rxni = ni;
4751         lnet_ni_addref_locked(ni, cpt);
4752         /* Multi-Rail: Primary NID of source. */
4753         msg->msg_initiator = lnet_peer_primary_nid_locked(src_nid);
4754
4755         /*
4756          * mark the status of this lpni as UP since we received a message
4757          * from it. The ping response reports back the ns_status which is
4758          * marked on the remote as up or down and we cache it here.
4759          */
4760         msg->msg_rxpeer->lpni_ns_status = LNET_NI_STATUS_UP;
4761
4762         lnet_msg_commit(msg, cpt);
4763
4764         /* message delay simulation */
4765         if (unlikely(!list_empty(&the_lnet.ln_delay_rules) &&
4766                      lnet_delay_rule_match_locked(hdr, msg))) {
4767                 lnet_net_unlock(cpt);
4768                 return 0;
4769         }
4770
4771         if (!for_me) {
4772                 rc = lnet_parse_forward_locked(ni, msg);
4773                 lnet_net_unlock(cpt);
4774
4775                 if (rc < 0)
4776                         goto free_drop;
4777
4778                 if (rc == LNET_CREDIT_OK) {
4779                         lnet_ni_recv(ni, msg->msg_private, msg, 0,
4780                                      0, payload_length, payload_length);
4781                 }
4782                 return 0;
4783         }
4784
4785         lnet_net_unlock(cpt);
4786
4787         rc = lnet_parse_local(ni, msg);
4788         if (rc != 0)
4789                 goto free_drop;
4790         return 0;
4791
4792  free_drop:
4793         LASSERT(msg->msg_md == NULL);
4794         lnet_finalize(msg, rc);
4795
4796  drop:
4797         lnet_drop_message(ni, cpt, private, payload_length, type);
4798         return 0;
4799 }
4800 EXPORT_SYMBOL(lnet_parse);
4801
4802 void
4803 lnet_drop_delayed_msg_list(struct list_head *head, char *reason)
4804 {
4805         while (!list_empty(head)) {
4806                 struct lnet_process_id id = {0};
4807                 struct lnet_msg *msg;
4808
4809                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4810                 list_del(&msg->msg_list);
4811
4812                 id.nid = msg->msg_hdr.src_nid;
4813                 id.pid = msg->msg_hdr.src_pid;
4814
4815                 LASSERT(msg->msg_md == NULL);
4816                 LASSERT(msg->msg_rx_delayed);
4817                 LASSERT(msg->msg_rxpeer != NULL);
4818                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4819
4820                 CWARN("Dropping delayed PUT from %s portal %d match %llu"
4821                       " offset %d length %d: %s\n",
4822                       libcfs_id2str(id),
4823                       msg->msg_hdr.msg.put.ptl_index,
4824                       msg->msg_hdr.msg.put.match_bits,
4825                       msg->msg_hdr.msg.put.offset,
4826                       msg->msg_hdr.payload_length, reason);
4827
4828                 /* NB I can't drop msg's ref on msg_rxpeer until after I've
4829                  * called lnet_drop_message(), so I just hang onto msg as well
4830                  * until that's done */
4831
4832                 lnet_drop_message(msg->msg_rxni, msg->msg_rx_cpt,
4833                                   msg->msg_private, msg->msg_len,
4834                                   msg->msg_type);
4835
4836                 msg->msg_no_resend = true;
4837                 /*
4838                  * NB: message will not generate event because w/o attached MD,
4839                  * but we still should give error code so lnet_msg_decommit()
4840                  * can skip counters operations and other checks.
4841                  */
4842                 lnet_finalize(msg, -ENOENT);
4843         }
4844 }
4845
4846 void
4847 lnet_recv_delayed_msg_list(struct list_head *head)
4848 {
4849         while (!list_empty(head)) {
4850                 struct lnet_msg *msg;
4851                 struct lnet_process_id id;
4852
4853                 msg = list_entry(head->next, struct lnet_msg, msg_list);
4854                 list_del(&msg->msg_list);
4855
4856                 /* md won't disappear under me, since each msg
4857                  * holds a ref on it */
4858
4859                 id.nid = msg->msg_hdr.src_nid;
4860                 id.pid = msg->msg_hdr.src_pid;
4861
4862                 LASSERT(msg->msg_rx_delayed);
4863                 LASSERT(msg->msg_md != NULL);
4864                 LASSERT(msg->msg_rxpeer != NULL);
4865                 LASSERT(msg->msg_rxni != NULL);
4866                 LASSERT(msg->msg_hdr.type == LNET_MSG_PUT);
4867
4868                 CDEBUG(D_NET, "Resuming delayed PUT from %s portal %d "
4869                        "match %llu offset %d length %d.\n",
4870                         libcfs_id2str(id), msg->msg_hdr.msg.put.ptl_index,
4871                         msg->msg_hdr.msg.put.match_bits,
4872                         msg->msg_hdr.msg.put.offset,
4873                         msg->msg_hdr.payload_length);
4874
4875                 lnet_recv_put(msg->msg_rxni, msg);
4876         }
4877 }
4878
4879 static void
4880 lnet_attach_rsp_tracker(struct lnet_rsp_tracker *rspt, int cpt,
4881                         struct lnet_libmd *md, struct lnet_handle_md mdh)
4882 {
4883         s64 timeout_ns;
4884         struct lnet_rsp_tracker *local_rspt;
4885
4886         /*
4887          * MD has a refcount taken by message so it's not going away.
4888          * The MD however can be looked up. We need to secure the access
4889          * to the md_rspt_ptr by taking the res_lock.
4890          * The rspt can be accessed without protection up to when it gets
4891          * added to the list.
4892          */
4893
4894         lnet_res_lock(cpt);
4895         local_rspt = md->md_rspt_ptr;
4896         timeout_ns = lnet_transaction_timeout * NSEC_PER_SEC;
4897         if (local_rspt != NULL) {
4898                 /*
4899                  * we already have an rspt attached to the md, so we'll
4900                  * update the deadline on that one.
4901                  */
4902                 lnet_rspt_free(rspt, cpt);
4903         } else {
4904                 /* new md */
4905                 rspt->rspt_mdh = mdh;
4906                 rspt->rspt_cpt = cpt;
4907                 /* store the rspt so we can access it when we get the REPLY */
4908                 md->md_rspt_ptr = rspt;
4909                 local_rspt = rspt;
4910         }
4911         local_rspt->rspt_deadline = ktime_add_ns(ktime_get(), timeout_ns);
4912
4913         /*
4914          * add to the list of tracked responses. It's added to tail of the
4915          * list in order to expire all the older entries first.
4916          */
4917         lnet_net_lock(cpt);
4918         list_move_tail(&local_rspt->rspt_on_list, the_lnet.ln_mt_rstq[cpt]);
4919         lnet_net_unlock(cpt);
4920         lnet_res_unlock(cpt);
4921 }
4922
4923 /**
4924  * Initiate an asynchronous PUT operation.
4925  *
4926  * There are several events associated with a PUT: completion of the send on
4927  * the initiator node (LNET_EVENT_SEND), and when the send completes
4928  * successfully, the receipt of an acknowledgment (LNET_EVENT_ACK) indicating
4929  * that the operation was accepted by the target. The event LNET_EVENT_PUT is
4930  * used at the target node to indicate the completion of incoming data
4931  * delivery.
4932  *
4933  * The local events will be logged in the EQ associated with the MD pointed to
4934  * by \a mdh handle. Using a MD without an associated EQ results in these
4935  * events being discarded. In this case, the caller must have another
4936  * mechanism (e.g., a higher level protocol) for determining when it is safe
4937  * to modify the memory region associated with the MD.
4938  *
4939  * Note that LNet does not guarantee the order of LNET_EVENT_SEND and
4940  * LNET_EVENT_ACK, though intuitively ACK should happen after SEND.
4941  *
4942  * \param self Indicates the NID of a local interface through which to send
4943  * the PUT request. Use LNET_NID_ANY to let LNet choose one by itself.
4944  * \param mdh A handle for the MD that describes the memory to be sent. The MD
4945  * must be "free floating" (See LNetMDBind()).
4946  * \param ack Controls whether an acknowledgment is requested.
4947  * Acknowledgments are only sent when they are requested by the initiating
4948  * process and the target MD enables them.
4949  * \param target A process identifier for the target process.
4950  * \param portal The index in the \a target's portal table.
4951  * \param match_bits The match bits to use for MD selection at the target
4952  * process.
4953  * \param offset The offset into the target MD (only used when the target
4954  * MD has the LNET_MD_MANAGE_REMOTE option set).
4955  * \param hdr_data 64 bits of user data that can be included in the message
4956  * header. This data is written to an event queue entry at the target if an
4957  * EQ is present on the matching MD.
4958  *
4959  * \retval  0      Success, and only in this case events will be generated
4960  * and logged to EQ (if it exists).
4961  * \retval -EIO    Simulated failure.
4962  * \retval -ENOMEM Memory allocation failure.
4963  * \retval -ENOENT Invalid MD object.
4964  *
4965  * \see struct lnet_event::hdr_data and lnet_event_kind_t.
4966  */
4967 int
4968 LNetPut(lnet_nid_t self, struct lnet_handle_md mdh, enum lnet_ack_req ack,
4969         struct lnet_process_id target, unsigned int portal,
4970         __u64 match_bits, unsigned int offset,
4971         __u64 hdr_data)
4972 {
4973         struct lnet_msg *msg;
4974         struct lnet_libmd *md;
4975         int cpt;
4976         int rc;
4977         struct lnet_rsp_tracker *rspt = NULL;
4978
4979         LASSERT(the_lnet.ln_refcount > 0);
4980
4981         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
4982             fail_peer(target.nid, 1)) {                 /* shall we now? */
4983                 CERROR("Dropping PUT to %s: simulated failure\n",
4984                        libcfs_id2str(target));
4985                 return -EIO;
4986         }
4987
4988         msg = lnet_msg_alloc();
4989         if (msg == NULL) {
4990                 CERROR("Dropping PUT to %s: ENOMEM on struct lnet_msg\n",
4991                        libcfs_id2str(target));
4992                 return -ENOMEM;
4993         }
4994         msg->msg_vmflush = !!(current->flags & PF_MEMALLOC);
4995
4996         cpt = lnet_cpt_of_cookie(mdh.cookie);
4997
4998         if (ack == LNET_ACK_REQ) {
4999                 rspt = lnet_rspt_alloc(cpt);
5000                 if (!rspt) {
5001                         CERROR("Dropping PUT to %s: ENOMEM on response tracker\n",
5002                                 libcfs_id2str(target));
5003                         return -ENOMEM;
5004                 }
5005                 INIT_LIST_HEAD(&rspt->rspt_on_list);
5006         }
5007
5008         lnet_res_lock(cpt);
5009
5010         md = lnet_handle2md(&mdh);
5011         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5012                 CERROR("Dropping PUT (%llu:%d:%s): MD (%d) invalid\n",
5013                        match_bits, portal, libcfs_id2str(target),
5014                        md == NULL ? -1 : md->md_threshold);
5015                 if (md != NULL && md->md_me != NULL)
5016                         CERROR("Source MD also attached to portal %d\n",
5017                                md->md_me->me_portal);
5018                 lnet_res_unlock(cpt);
5019
5020                 if (rspt)
5021                         lnet_rspt_free(rspt, cpt);
5022
5023                 lnet_msg_free(msg);
5024                 return -ENOENT;
5025         }
5026
5027         CDEBUG(D_NET, "LNetPut -> %s\n", libcfs_id2str(target));
5028
5029         lnet_msg_attach_md(msg, md, 0, 0);
5030
5031         lnet_prep_send(msg, LNET_MSG_PUT, target, 0, md->md_length);
5032
5033         msg->msg_hdr.msg.put.match_bits = cpu_to_le64(match_bits);
5034         msg->msg_hdr.msg.put.ptl_index = cpu_to_le32(portal);
5035         msg->msg_hdr.msg.put.offset = cpu_to_le32(offset);
5036         msg->msg_hdr.msg.put.hdr_data = hdr_data;
5037
5038         /* NB handles only looked up by creator (no flips) */
5039         if (ack == LNET_ACK_REQ) {
5040                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5041                         the_lnet.ln_interface_cookie;
5042                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5043                         md->md_lh.lh_cookie;
5044         } else {
5045                 msg->msg_hdr.msg.put.ack_wmd.wh_interface_cookie =
5046                         LNET_WIRE_HANDLE_COOKIE_NONE;
5047                 msg->msg_hdr.msg.put.ack_wmd.wh_object_cookie =
5048                         LNET_WIRE_HANDLE_COOKIE_NONE;
5049         }
5050
5051         lnet_res_unlock(cpt);
5052
5053         lnet_build_msg_event(msg, LNET_EVENT_SEND);
5054
5055         if (rspt && lnet_response_tracking_enabled(LNET_MSG_PUT,
5056                                                    md->md_options))
5057                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5058         else if (rspt)
5059                 lnet_rspt_free(rspt, cpt);
5060
5061         if (CFS_FAIL_CHECK_ORSET(CFS_FAIL_PTLRPC_OST_BULK_CB2,
5062                                  CFS_FAIL_ONCE))
5063                 rc = -EIO;
5064         else
5065                 rc = lnet_send(self, msg, LNET_NID_ANY);
5066
5067         if (rc != 0) {
5068                 CNETERR("Error sending PUT to %s: %d\n",
5069                         libcfs_id2str(target), rc);
5070                 msg->msg_no_resend = true;
5071                 lnet_finalize(msg, rc);
5072         }
5073
5074         /* completion will be signalled by an event */
5075         return 0;
5076 }
5077 EXPORT_SYMBOL(LNetPut);
5078
5079 /*
5080  * The LND can DMA direct to the GET md (i.e. no REPLY msg).  This
5081  * returns a msg for the LND to pass to lnet_finalize() when the sink
5082  * data has been received.
5083  *
5084  * CAVEAT EMPTOR: 'getmsg' is the original GET, which is freed when
5085  * lnet_finalize() is called on it, so the LND must call this first
5086  */
5087 struct lnet_msg *
5088 lnet_create_reply_msg(struct lnet_ni *ni, struct lnet_msg *getmsg)
5089 {
5090         struct lnet_msg *msg = lnet_msg_alloc();
5091         struct lnet_libmd *getmd = getmsg->msg_md;
5092         struct lnet_process_id peer_id = getmsg->msg_target;
5093         int cpt;
5094
5095         LASSERT(!getmsg->msg_target_is_router);
5096         LASSERT(!getmsg->msg_routing);
5097
5098         if (msg == NULL) {
5099                 CERROR("%s: Dropping REPLY from %s: can't allocate msg\n",
5100                        libcfs_nidstr(&ni->ni_nid), libcfs_id2str(peer_id));
5101                 goto drop;
5102         }
5103
5104         cpt = lnet_cpt_of_cookie(getmd->md_lh.lh_cookie);
5105         lnet_res_lock(cpt);
5106
5107         LASSERT(getmd->md_refcount > 0);
5108
5109         if (getmd->md_threshold == 0) {
5110                 CERROR("%s: Dropping REPLY from %s for inactive MD %p\n",
5111                         libcfs_nidstr(&ni->ni_nid), libcfs_id2str(peer_id),
5112                         getmd);
5113                 lnet_res_unlock(cpt);
5114                 goto drop;
5115         }
5116
5117         LASSERT(getmd->md_offset == 0);
5118
5119         CDEBUG(D_NET, "%s: Reply from %s md %p\n",
5120                libcfs_nidstr(&ni->ni_nid), libcfs_id2str(peer_id), getmd);
5121
5122         /* setup information for lnet_build_msg_event */
5123         msg->msg_initiator =
5124                 lnet_nid_to_nid4(&getmsg->msg_txpeer->lpni_peer_net->lpn_peer->lp_primary_nid);
5125         msg->msg_from = peer_id.nid;
5126         msg->msg_type = LNET_MSG_GET; /* flag this msg as an "optimized" GET */
5127         msg->msg_hdr.src_nid = peer_id.nid;
5128         msg->msg_hdr.payload_length = getmd->md_length;
5129         msg->msg_receiving = 1; /* required by lnet_msg_attach_md */
5130
5131         lnet_msg_attach_md(msg, getmd, getmd->md_offset, getmd->md_length);
5132         lnet_res_unlock(cpt);
5133
5134         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
5135
5136         lnet_net_lock(cpt);
5137         lnet_msg_commit(msg, cpt);
5138         lnet_net_unlock(cpt);
5139
5140         lnet_build_msg_event(msg, LNET_EVENT_REPLY);
5141
5142         return msg;
5143
5144  drop:
5145         cpt = lnet_cpt_of_nid(peer_id.nid, ni);
5146
5147         lnet_net_lock(cpt);
5148         lnet_incr_stats(&ni->ni_stats, LNET_MSG_GET, LNET_STATS_TYPE_DROP);
5149         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_count++;
5150         the_lnet.ln_counters[cpt]->lct_common.lcc_drop_length +=
5151                 getmd->md_length;
5152         lnet_net_unlock(cpt);
5153
5154         if (msg != NULL)
5155                 lnet_msg_free(msg);
5156
5157         return NULL;
5158 }
5159 EXPORT_SYMBOL(lnet_create_reply_msg);
5160
5161 void
5162 lnet_set_reply_msg_len(struct lnet_ni *ni, struct lnet_msg *reply,
5163                        unsigned int len)
5164 {
5165         /* Set the REPLY length, now the RDMA that elides the REPLY message has
5166          * completed and I know it. */
5167         LASSERT(reply != NULL);
5168         LASSERT(reply->msg_type == LNET_MSG_GET);
5169         LASSERT(reply->msg_ev.type == LNET_EVENT_REPLY);
5170
5171         /* NB I trusted my peer to RDMA.  If she tells me she's written beyond
5172          * the end of my buffer, I might as well be dead. */
5173         LASSERT(len <= reply->msg_ev.mlength);
5174
5175         reply->msg_ev.mlength = len;
5176 }
5177 EXPORT_SYMBOL(lnet_set_reply_msg_len);
5178
5179 /**
5180  * Initiate an asynchronous GET operation.
5181  *
5182  * On the initiator node, an LNET_EVENT_SEND is logged when the GET request
5183  * is sent, and an LNET_EVENT_REPLY is logged when the data returned from
5184  * the target node in the REPLY has been written to local MD.
5185  *
5186  * On the target node, an LNET_EVENT_GET is logged when the GET request
5187  * arrives and is accepted into a MD.
5188  *
5189  * \param self,target,portal,match_bits,offset See the discussion in LNetPut().
5190  * \param mdh A handle for the MD that describes the memory into which the
5191  * requested data will be received. The MD must be "free floating" (See LNetMDBind()).
5192  *
5193  * \retval  0      Success, and only in this case events will be generated
5194  * and logged to EQ (if it exists) of the MD.
5195  * \retval -EIO    Simulated failure.
5196  * \retval -ENOMEM Memory allocation failure.
5197  * \retval -ENOENT Invalid MD object.
5198  */
5199 int
5200 LNetGet(lnet_nid_t self, struct lnet_handle_md mdh,
5201         struct lnet_process_id target, unsigned int portal,
5202         __u64 match_bits, unsigned int offset, bool recovery)
5203 {
5204         struct lnet_msg *msg;
5205         struct lnet_libmd *md;
5206         struct lnet_rsp_tracker *rspt;
5207         int cpt;
5208         int rc;
5209
5210         LASSERT(the_lnet.ln_refcount > 0);
5211
5212         if (!list_empty(&the_lnet.ln_test_peers) &&     /* normally we don't */
5213             fail_peer(target.nid, 1))                   /* shall we now? */
5214         {
5215                 CERROR("Dropping GET to %s: simulated failure\n",
5216                        libcfs_id2str(target));
5217                 return -EIO;
5218         }
5219
5220         msg = lnet_msg_alloc();
5221         if (!msg) {
5222                 CERROR("Dropping GET to %s: ENOMEM on struct lnet_msg\n",
5223                        libcfs_id2str(target));
5224                 return -ENOMEM;
5225         }
5226
5227         cpt = lnet_cpt_of_cookie(mdh.cookie);
5228
5229         rspt = lnet_rspt_alloc(cpt);
5230         if (!rspt) {
5231                 CERROR("Dropping GET to %s: ENOMEM on response tracker\n",
5232                        libcfs_id2str(target));
5233                 return -ENOMEM;
5234         }
5235         INIT_LIST_HEAD(&rspt->rspt_on_list);
5236
5237         msg->msg_recovery = recovery;
5238
5239         lnet_res_lock(cpt);
5240
5241         md = lnet_handle2md(&mdh);
5242         if (md == NULL || md->md_threshold == 0 || md->md_me != NULL) {
5243                 CERROR("Dropping GET (%llu:%d:%s): MD (%d) invalid\n",
5244                        match_bits, portal, libcfs_id2str(target),
5245                        md == NULL ? -1 : md->md_threshold);
5246                 if (md != NULL && md->md_me != NULL)
5247                         CERROR("REPLY MD also attached to portal %d\n",
5248                                md->md_me->me_portal);
5249
5250                 lnet_res_unlock(cpt);
5251
5252                 lnet_msg_free(msg);
5253                 lnet_rspt_free(rspt, cpt);
5254                 return -ENOENT;
5255         }
5256
5257         CDEBUG(D_NET, "LNetGet -> %s\n", libcfs_id2str(target));
5258
5259         lnet_msg_attach_md(msg, md, 0, 0);
5260
5261         lnet_prep_send(msg, LNET_MSG_GET, target, 0, 0);
5262
5263         msg->msg_hdr.msg.get.match_bits = cpu_to_le64(match_bits);
5264         msg->msg_hdr.msg.get.ptl_index = cpu_to_le32(portal);
5265         msg->msg_hdr.msg.get.src_offset = cpu_to_le32(offset);
5266         msg->msg_hdr.msg.get.sink_length = cpu_to_le32(md->md_length);
5267
5268         /* NB handles only looked up by creator (no flips) */
5269         msg->msg_hdr.msg.get.return_wmd.wh_interface_cookie =
5270                 the_lnet.ln_interface_cookie;
5271         msg->msg_hdr.msg.get.return_wmd.wh_object_cookie =
5272                 md->md_lh.lh_cookie;
5273
5274         lnet_res_unlock(cpt);
5275
5276         lnet_build_msg_event(msg, LNET_EVENT_SEND);
5277
5278         if (lnet_response_tracking_enabled(LNET_MSG_GET, md->md_options))
5279                 lnet_attach_rsp_tracker(rspt, cpt, md, mdh);
5280         else
5281                 lnet_rspt_free(rspt, cpt);
5282
5283         rc = lnet_send(self, msg, LNET_NID_ANY);
5284         if (rc < 0) {
5285                 CNETERR("Error sending GET to %s: %d\n",
5286                         libcfs_id2str(target), rc);
5287                 msg->msg_no_resend = true;
5288                 lnet_finalize(msg, rc);
5289         }
5290
5291         /* completion will be signalled by an event */
5292         return 0;
5293 }
5294 EXPORT_SYMBOL(LNetGet);
5295
5296 /**
5297  * Calculate distance to node at \a dstnid.
5298  *
5299  * \param dstnid Target NID.
5300  * \param srcnidp If not NULL, NID of the local interface to reach \a dstnid
5301  * is saved here.
5302  * \param orderp If not NULL, order of the route to reach \a dstnid is saved
5303  * here.
5304  *
5305  * \retval 0 If \a dstnid belongs to a local interface, and reserved option
5306  * local_nid_dist_zero is set, which is the default.
5307  * \retval positives Distance to target NID, i.e. number of hops plus one.
5308  * \retval -EHOSTUNREACH If \a dstnid is not reachable.
5309  */
5310 int
5311 LNetDist(lnet_nid_t dstnid, lnet_nid_t *srcnidp, __u32 *orderp)
5312 {
5313         struct list_head *e;
5314         struct lnet_ni *ni = NULL;
5315         struct lnet_remotenet *rnet;
5316         __u32 dstnet = LNET_NIDNET(dstnid);
5317         int hops;
5318         int cpt;
5319         __u32 order = 2;
5320         struct list_head *rn_list;
5321         bool matched_dstnet = false;
5322
5323         /* if !local_nid_dist_zero, I don't return a distance of 0 ever
5324          * (when lustre sees a distance of 0, it substitutes 0@lo), so I
5325          * keep order 0 free for 0@lo and order 1 free for a local NID
5326          * match */
5327
5328         LASSERT(the_lnet.ln_refcount > 0);
5329
5330         cpt = lnet_net_lock_current();
5331
5332         while ((ni = lnet_get_next_ni_locked(NULL, ni))) {
5333                 /* FIXME support large-addr nid */
5334                 if (lnet_nid_to_nid4(&ni->ni_nid) == dstnid) {
5335                         if (srcnidp != NULL)
5336                                 *srcnidp = dstnid;
5337                         if (orderp != NULL) {
5338                                 if (dstnid == LNET_NID_LO_0)
5339                                         *orderp = 0;
5340                                 else
5341                                         *orderp = 1;
5342                         }
5343                         lnet_net_unlock(cpt);
5344
5345                         return local_nid_dist_zero ? 0 : 1;
5346                 }
5347
5348                 if (!matched_dstnet && LNET_NID_NET(&ni->ni_nid) == dstnet) {
5349                         matched_dstnet = true;
5350                         /* We matched the destination net, but we may have
5351                          * additional local NIs to inspect.
5352                          *
5353                          * We record the nid and order as appropriate, but
5354                          * they may be overwritten if we match local NI above.
5355                          */
5356                         if (srcnidp)
5357                                 /* FIXME support large-addr nids */
5358                                 *srcnidp = lnet_nid_to_nid4(&ni->ni_nid);
5359
5360                         if (orderp) {
5361                                 /* Check if ni was originally created in
5362                                  * current net namespace.
5363                                  * If not, assign order above 0xffff0000,
5364                                  * to make this ni not a priority.
5365                                  */
5366                                 if (current->nsproxy &&
5367                                     !net_eq(ni->ni_net_ns,
5368                                             current->nsproxy->net_ns))
5369                                         *orderp = order + 0xffff0000;
5370                                 else
5371                                         *orderp = order;
5372                         }
5373                 }
5374
5375                 order++;
5376         }
5377
5378         if (matched_dstnet) {
5379                 lnet_net_unlock(cpt);
5380                 return 1;
5381         }
5382
5383         rn_list = lnet_net2rnethash(dstnet);
5384         list_for_each(e, rn_list) {
5385                 rnet = list_entry(e, struct lnet_remotenet, lrn_list);
5386
5387                 if (rnet->lrn_net == dstnet) {
5388                         struct lnet_route *route;
5389                         struct lnet_route *shortest = NULL;
5390                         __u32 shortest_hops = LNET_UNDEFINED_HOPS;
5391                         __u32 route_hops;
5392
5393                         LASSERT(!list_empty(&rnet->lrn_routes));
5394
5395                         list_for_each_entry(route, &rnet->lrn_routes,
5396                                             lr_list) {
5397                                 route_hops = route->lr_hops;
5398                                 if (route_hops == LNET_UNDEFINED_HOPS)
5399                                         route_hops = 1;
5400                                 if (shortest == NULL ||
5401                                     route_hops < shortest_hops) {
5402                                         shortest = route;
5403                                         shortest_hops = route_hops;
5404                                 }
5405                         }
5406
5407                         LASSERT(shortest != NULL);
5408                         hops = shortest_hops;
5409                         if (srcnidp != NULL) {
5410                                 struct lnet_net *net;
5411                                 net = lnet_get_net_locked(shortest->lr_lnet);
5412                                 LASSERT(net);
5413                                 ni = lnet_get_next_ni_locked(net, NULL);
5414                                 /* FIXME support large-addr nids */
5415                                 *srcnidp = lnet_nid_to_nid4(&ni->ni_nid);
5416                         }
5417                         if (orderp != NULL)
5418                                 *orderp = order;
5419                         lnet_net_unlock(cpt);
5420                         return hops + 1;
5421                 }
5422                 order++;
5423         }
5424
5425         lnet_net_unlock(cpt);
5426         return -EHOSTUNREACH;
5427 }
5428 EXPORT_SYMBOL(LNetDist);